diff --git a/OCL20ToSPIRV.cpp b/OCL20ToSPIRV.cpp
index 1a052e4..8606d60 100644
--- a/OCL20ToSPIRV.cpp
+++ b/OCL20ToSPIRV.cpp
@@ -1,1457 +1,1457 @@
-//===- OCL20ToSPIRV.cpp - Transform OCL20 to SPIR-V builtins -----*- C++ -*-===//
-//
-//                     The LLVM/SPIRV Translator
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-// Copyright (c) 2014 Advanced Micro Devices, Inc. All rights reserved.
-//
-// Permission is hereby granted, free of charge, to any person obtaining a
-// copy of this software and associated documentation files (the "Software"),
-// to deal with the Software without restriction, including without limitation
-// the rights to use, copy, modify, merge, publish, distribute, sublicense,
-// and/or sell copies of the Software, and to permit persons to whom the
-// Software is furnished to do so, subject to the following conditions:
-//
-// Redistributions of source code must retain the above copyright notice,
-// this list of conditions and the following disclaimers.
-// Redistributions in binary form must reproduce the above copyright notice,
-// this list of conditions and the following disclaimers in the documentation
-// and/or other materials provided with the distribution.
-// Neither the names of Advanced Micro Devices, Inc., nor the names of its
-// contributors may be used to endorse or promote products derived from this
-// Software without specific prior written permission.
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-// CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
-// THE SOFTWARE.
-//
-//===----------------------------------------------------------------------===//
-//
-// This file implements translation of OCL20 builtin functions.
-//
-//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "cl20tospv"
-
-#include "SPIRVInternal.h"
-#include "OCLUtil.h"
-#include "OCLTypeToSPIRV.h"
-
-#include "llvm/ADT/StringSwitch.h"
-#include "llvm/IR/InstVisitor.h"
-#include "llvm/IR/Instructions.h"
-#include "llvm/IR/Instruction.h"
-#include "llvm/IR/IRBuilder.h"
-#include "llvm/IR/Verifier.h"
-#include "llvm/Pass.h"
-#include "llvm/PassSupport.h"
-#include "llvm/Support/Debug.h"
-#include "llvm/Support/raw_ostream.h"
-
-#include <set>
-
-using namespace llvm;
-using namespace SPIRV;
-using namespace OCLUtil;
-
-namespace SPIRV {
-static size_t
-getOCLCpp11AtomicMaxNumOps(StringRef Name) {
-  return StringSwitch<size_t>(Name)
-      .Cases("load", "flag_test_and_set", "flag_clear", 3)
-      .Cases("store", "exchange",  4)
-      .StartsWith("compare_exchange", 6)
-      .StartsWith("fetch", 4)
-      .Default(0);
-}
-
-class OCL20ToSPIRV: public ModulePass,
-  public InstVisitor<OCL20ToSPIRV> {
-public:
-  OCL20ToSPIRV():ModulePass(ID), M(nullptr), Ctx(nullptr), CLVer(0) {
-    initializeOCL20ToSPIRVPass(*PassRegistry::getPassRegistry());
-  }
-  virtual bool runOnModule(Module &M);
-
-  void getAnalysisUsage(AnalysisUsage &AU) const {
-    AU.addRequired<OCLTypeToSPIRV>();
-  }
-
-  virtual void visitCallInst(CallInst &CI);
-
-  /// Transform barrier/work_group_barrier to __spirv_ControlBarrier.
-  /// barrier(flag) =>
-  ///   __spirv_ControlBarrier(workgroup, workgroup, map(flag))
-  /// workgroup_barrier(scope, flag) =>
-  ///   __spirv_ControlBarrier(workgroup, map(scope), map(flag))
-  void visitCallWorkGroupBarrier(CallInst *CI);
-
-  /// Erase useless convert functions.
-  /// \return true if the call instruction is erased.
-  bool eraseUselessConvert(CallInst *Call, const std::string &MangledName,
-      const std::string &DeMangledName);
-
-  /// Transform convert_ to
-  ///   __spirv_{CastOpName}_R{TargeTyName}{_sat}{_rt[p|n|z|e]}
-  void visitCallConvert(CallInst *CI, StringRef MangledName,
-    const std::string &DemangledName);
-
-  /// Transform async_work_group{_strided}_copy.
-  /// async_work_group_copy(dst, src, n, event)
-  ///   => async_work_group_strided_copy(dst, src, n, 1, event)
-  /// async_work_group_strided_copy(dst, src, n, stride, event)
-  ///   => __spirv_AsyncGroupCopy(ScopeWorkGroup, dst, src, n, stride, event)
-  void visitCallAsyncWorkGroupCopy(CallInst *CI,
-      const std::string &DemangledName);
-
-  /// Transform OCL builtin function to SPIR-V builtin function.
-  void transBuiltin(CallInst *CI, OCLBuiltinTransInfo &Info);
-
-  /// Transform OCL work item builtin functions to SPIR-V builtin variables.
-  void transWorkItemBuiltinsToVariables();
-
-  /// Transform atomic_work_item_fence/mem_fence to __spirv_MemoryBarrier.
-  /// func(flag, order, scope) =>
-  ///   __spirv_MemoryBarrier(map(scope), map(flag)|map(order))
-  void transMemoryBarrier(CallInst *CI, AtomicWorkItemFenceLiterals);
-
-  /// Transform all to __spirv_Op(All|Any).  Note that the types mismatch so
-  // some extra code is emitted to convert between the two.
-  void visitCallAllAny(spv::Op OC, CallInst *CI);
-
-  /// Transform atomic_* to __spirv_Atomic*.
-  /// atomic_x(ptr_arg, args, order, scope) =>
-  ///   __spirv_AtomicY(ptr_arg, map(order), map(scope), args)
-  void transAtomicBuiltin(CallInst *CI, OCLBuiltinTransInfo &Info);
-
-  /// Transform atomic_work_item_fence to __spirv_MemoryBarrier.
-  /// atomic_work_item_fence(flag, order, scope) =>
-  ///   __spirv_MemoryBarrier(map(scope), map(flag)|map(order))
-  void visitCallAtomicWorkItemFence(CallInst *CI);
-
-  /// Transform atomic_compare_exchange call.
-  /// In atomic_compare_exchange, the expected value parameter is a pointer.
-  /// However in SPIR-V it is a value. The transformation adds a load 
-  /// instruction, result of which is passed to atomic_compare_exchange as
-  /// argument.
-  /// The transformation adds a store instruction after the call, to update the
-  /// value in expected with the value pointed to by object. Though, it is not
-  /// necessary in case they are equal, this approach makes result code simpler.
-  /// Also ICmp instruction is added, because the call must return result of
-  /// comparison.
-  /// \returns the call instruction of atomic_compare_exchange_strong.
-  CallInst *visitCallAtomicCmpXchg(CallInst *CI,
-      const std::string &DemangledName);
-
-  /// Transform atomic_init.
-  /// atomic_init(p, x) => store p, x
-  void visitCallAtomicInit(CallInst *CI);
-
-  /// Transform legacy OCL 1.x atomic builtins to SPIR-V builtins for extensions
-  ///   cl_khr_int64_base_atomics
-  ///   cl_khr_int64_extended_atomics
-  /// Do nothing if the called function is not a legacy atomic builtin.
-  void visitCallAtomicLegacy(CallInst *CI, StringRef MangledName,
-    const std::string &DemangledName);
-
-  /// Transform OCL 2.0 C++11 atomic builtins to SPIR-V builtins.
-  /// Do nothing if the called function is not a C++11 atomic builtin.
-  void visitCallAtomicCpp11(CallInst *CI, StringRef MangledName,
-    const std::string &DemangledName);
-
-  /// Transform OCL builtin function to SPIR-V builtin function.
-  /// Assuming there is a simple name mapping without argument changes.
-  /// Should be called at last.
-  void visitCallBuiltinSimple(CallInst *CI, StringRef MangledName,
-    const std::string &DemangledName);
-
-  /// Transform get_image_{width|height|depth|dim}.
-  /// get_image_xxx(...) =>
-  ///   dimension = __spirv_ImageQuerySizeLod_R{ReturnType}(...);
-  ///   return dimension.{x|y|z};
-  void visitCallGetImageSize(CallInst *CI, StringRef MangledName,
-    const std::string &DemangledName);
-
-  /// Transform {work|sub}_group_x =>
-  ///   __spirv_{OpName}
-  ///
-  /// Special handling of work_group_broadcast.
-  ///   work_group_broadcast(a, x, y, z)
-  ///     =>
-  ///   __spirv_GroupBroadcast(a, vec3(x, y, z))
-
-  void visitCallGroupBuiltin(CallInst *CI, StringRef MangledName,
-    const std::string &DemangledName);
-
-  /// Transform mem_fence to __spirv_MemoryBarrier.
-  /// mem_fence(flag) => __spirv_MemoryBarrier(Workgroup, map(flag))
-  void visitCallMemFence(CallInst *CI);
-
-  void visitCallNDRange(CallInst *CI, const std::string &DemangledName);
-
-  /// Transform OCL pipe builtin function to SPIR-V pipe builtin function.
-  void visitCallPipeBuiltin(CallInst *CI, StringRef MangledName,
-    const std::string &DemangledName);
-
-  /// Transform read_image with sampler arguments.
-  /// read_image(image, sampler, ...) =>
-  ///   sampled_image = __spirv_SampledImage(image, sampler);
-  ///   return __spirv_ImageSampleExplicitLod_R{ReturnType}(sampled_image, ...);
-  void visitCallReadImageWithSampler(CallInst *CI, StringRef MangledName,
-      const std::string &DemangledName);
-
-  /// Transform read_image with msaa image arguments.
-  /// Sample argument must be acoded as Image Operand.
-  void visitCallReadImageMSAA(CallInst *CI, StringRef MangledName,
-                              const std::string &DemangledName);
-
-  /// Transform {read|write}_image without sampler arguments.
-  void visitCallReadWriteImage(CallInst *CI, StringRef MangledName,
-      const std::string &DemangledName);
-
-  /// Transform to_{global|local|private}.
-  ///
-  /// T* a = ...;
-  /// addr T* b = to_addr(a);
-  ///   =>
-  /// i8* x = cast<i8*>(a);
-  /// addr i8* y = __spirv_GenericCastToPtr_ToAddr(x);
-  /// addr T* b = cast<addr T*>(y);
-  void visitCallToAddr(CallInst *CI, StringRef MangledName,
-      const std::string &DemangledName);
-
-  /// Transform return type of relatinal built-in functions like isnan, isfinite
-  /// to boolean values.
-  void visitCallRelational(CallInst *CI, const std::string &DemangledName);
-
-  /// Transform vector load/store functions to SPIR-V extended builtin
-  ///   functions
-  /// {vload|vstore{a}}{_half}{n}{_rte|_rtz|_rtp|_rtn} =>
-  ///   __spirv_ocl_{ExtendedInstructionOpCodeName}__R{ReturnType}
-  void visitCallVecLoadStore(CallInst *CI, StringRef MangledName,
-      const std::string &DemangledName);
-
-  /// Transforms get_mem_fence built-in to SPIR-V function and aligns result values with SPIR 1.2.
-  /// get_mem_fence(ptr) => __spirv_GenericPtrMemSemantics
-  /// GenericPtrMemSemantics valid values are 0x100, 0x200 and 0x300, where is
-  /// SPIR 1.2 defines them as 0x1, 0x2 and 0x3, so this function adjusts
-  /// GenericPtrMemSemantics results to SPIR 1.2 values.
-  void visitCallGetFence(CallInst *CI, StringRef MangledName, const std::string& DemangledName);
-
-  /// Transforms OpDot instructions with a scalar type to a fmul instruction
-  void visitCallDot(CallInst *CI);
-
-  /// Fixes for built-in functions with vector+scalar arguments that are
-  /// translated to the SPIR-V instructions where all arguments must have the
-  /// same type.
-  void visitCallScalToVec(CallInst *CI, StringRef MangledName,
-                          const std::string &DemangledName);
-
-  /// Transform get_image_channel_{order|data_type} built-in functions to
-  ///   __spirv_ocl_{ImageQueryOrder|ImageQueryFormat}
-  void visitCallGetImageChannel(CallInst *CI, StringRef MangledName,
-                                const std::string &DemangledName,
-                                unsigned int Offset);
-
-  void visitDbgInfoIntrinsic(DbgInfoIntrinsic &I){
-    I.dropAllReferences();
-    I.eraseFromParent();
-  }
-  static char ID;
-private:
-  Module *M;
-  LLVMContext *Ctx;
-  unsigned CLVer;                   /// OpenCL version as major*10+minor
-  std::set<Value *> ValuesToDelete;
-
-  ConstantInt *addInt32(int I) {
-    return getInt32(M, I);
-  }
-  ConstantInt *addSizet(uint64_t I) {
-    return getSizet(M, I);
-  }
-
-  /// Get vector width from OpenCL vload* function name.
-  SPIRVWord getVecLoadWidth(const std::string& DemangledName) {
-    SPIRVWord Width = 0;
-    if (DemangledName == "vloada_half")
-      Width = 1;
-    else {
-      unsigned Loc = 5;
-      if (DemangledName.find("vload_half") == 0)
-        Loc = 10;
-      else if (DemangledName.find("vloada_half") == 0)
-        Loc = 11;
-
-      std::stringstream SS(DemangledName.substr(Loc));
-      SS >> Width;
-    }
-    return Width;
-  }
-
-  /// Transform OpenCL vload/vstore function name.
-  void transVecLoadStoreName(std::string& DemangledName,
-      const std::string &Stem, bool AlwaysN) {
-    auto HalfStem = Stem + "_half";
-    auto HalfStemR = HalfStem + "_r";
-    if (!AlwaysN && DemangledName == HalfStem)
-      return;
-    if (!AlwaysN && DemangledName.find(HalfStemR) == 0) {
-      DemangledName = HalfStemR;
-      return;
-    }
-    if (DemangledName.find(HalfStem) == 0) {
-      auto OldName = DemangledName;
-      DemangledName = HalfStem + "n";
-      if (OldName.find("_r") != std::string::npos)
-        DemangledName += "_r";
-      return;
-    }
-    if (DemangledName.find(Stem) == 0) {
-      DemangledName = Stem + "n";
-      return;
-    }
-  }
-
-};
-
-char OCL20ToSPIRV::ID = 0;
-
-bool
-OCL20ToSPIRV::runOnModule(Module& Module) {
-  M = &Module;
-  Ctx = &M->getContext();
-  auto Src = getSPIRVSource(&Module);
-  if (std::get<0>(Src) != spv::SourceLanguageOpenCL_C)
-    return false;
-
-  CLVer = std::get<1>(Src);
-  if (CLVer > kOCLVer::CL20)
-    return false;
-
-  DEBUG(dbgs() << "Enter OCL20ToSPIRV:\n");
-
-  transWorkItemBuiltinsToVariables();
-
-  visit(*M);
-
-  for (auto &I:ValuesToDelete)
-    if (auto Inst = dyn_cast<Instruction>(I))
-      Inst->eraseFromParent();
-  for (auto &I:ValuesToDelete)
-    if (auto GV = dyn_cast<GlobalValue>(I))
-      GV->eraseFromParent();
-
-  DEBUG(dbgs() << "After OCL20ToSPIRV:\n" << *M);
-
-  std::string Err;
-  raw_string_ostream ErrorOS(Err);
-  if (verifyModule(*M, &ErrorOS)){
-    DEBUG(errs() << "Fails to verify module: " << ErrorOS.str());
-  }
-  return true;
-}
-
-// The order of handling OCL builtin functions is important.
-// Workgroup functions need to be handled before pipe functions since
-// there are functions fall into both categories.
-void
-OCL20ToSPIRV::visitCallInst(CallInst& CI) {
-  DEBUG(dbgs() << "[visistCallInst] " << CI << '\n');
-  auto F = CI.getCalledFunction();
-  if (!F)
-    return;
-
-  auto MangledName = F->getName();
-  std::string DemangledName;
-  if (!oclIsBuiltin(MangledName, &DemangledName))
-    return;
-  DEBUG(dbgs() << "DemangledName: " << DemangledName << '\n');
-  if (DemangledName.find(kOCLBuiltinName::NDRangePrefix) == 0) {
-    visitCallNDRange(&CI, DemangledName);
-    return;
-  }
-  if (DemangledName == kOCLBuiltinName::All) {
-      visitCallAllAny(OpAll, &CI);
-      return;
-  }
-  if (DemangledName == kOCLBuiltinName::Any) {
-      visitCallAllAny(OpAny, &CI);
-      return;
-  }
-  if (DemangledName.find(kOCLBuiltinName::AsyncWorkGroupCopy) == 0 ||
-      DemangledName.find(kOCLBuiltinName::AsyncWorkGroupStridedCopy) == 0) {
-    visitCallAsyncWorkGroupCopy(&CI, DemangledName);
-    return;
-  }
-  if (DemangledName.find(kOCLBuiltinName::AtomicPrefix) == 0 ||
-      DemangledName.find(kOCLBuiltinName::AtomPrefix) == 0) {
-    auto PCI = &CI;
-    if (DemangledName == kOCLBuiltinName::AtomicInit) {
-      visitCallAtomicInit(PCI);
-      return;
-    }
-    if (DemangledName == kOCLBuiltinName::AtomicWorkItemFence) {
-      visitCallAtomicWorkItemFence(PCI);
-      return;
-    }
-    if (DemangledName == kOCLBuiltinName::AtomicCmpXchgWeak ||
-        DemangledName == kOCLBuiltinName::AtomicCmpXchgStrong ||
-        DemangledName == kOCLBuiltinName::AtomicCmpXchgWeakExplicit ||
-        DemangledName == kOCLBuiltinName::AtomicCmpXchgStrongExplicit) {
-      assert(CLVer == kOCLVer::CL20 && "Wrong version of OpenCL");
-      PCI = visitCallAtomicCmpXchg(PCI, DemangledName);
-    }
-    visitCallAtomicLegacy(PCI, MangledName, DemangledName);
-    visitCallAtomicCpp11(PCI, MangledName, DemangledName);
-    return;
-  }
-  if (DemangledName.find(kOCLBuiltinName::ConvertPrefix) == 0) {
-    visitCallConvert(&CI, MangledName, DemangledName);
-    return;
-  }
-  if (DemangledName == kOCLBuiltinName::GetImageWidth ||
-      DemangledName == kOCLBuiltinName::GetImageHeight ||
-      DemangledName == kOCLBuiltinName::GetImageDepth ||
-      DemangledName == kOCLBuiltinName::GetImageDim   ||
-      DemangledName == kOCLBuiltinName::GetImageArraySize) {
-    visitCallGetImageSize(&CI, MangledName, DemangledName);
-    return;
-  }
-  if ((DemangledName.find(kOCLBuiltinName::WorkGroupPrefix) == 0 &&
-      DemangledName != kOCLBuiltinName::WorkGroupBarrier) ||
-      DemangledName == kOCLBuiltinName::WaitGroupEvent ||
-      DemangledName.find(kOCLBuiltinName::SubGroupPrefix) == 0) {
-    visitCallGroupBuiltin(&CI, MangledName, DemangledName);
-    return;
-  }
-  if (DemangledName.find(kOCLBuiltinName::Pipe) != std::string::npos) {
-    visitCallPipeBuiltin(&CI, MangledName, DemangledName);
-    return;
-  }
-  if (DemangledName == kOCLBuiltinName::MemFence) {
-    visitCallMemFence(&CI);
-    return;
-  }
-  if (DemangledName.find(kOCLBuiltinName::ReadImage) == 0) {
-    if (MangledName.find(kMangledName::Sampler) != StringRef::npos) {
-      visitCallReadImageWithSampler(&CI, MangledName, DemangledName);
-      return;
-    }
-    if (MangledName.find("msaa") != StringRef::npos) {
-      visitCallReadImageMSAA(&CI, MangledName, DemangledName);
-      return;
-    }
-  }
-  if (DemangledName.find(kOCLBuiltinName::ReadImage) == 0 ||
-      DemangledName.find(kOCLBuiltinName::WriteImage) == 0) {
-    visitCallReadWriteImage(&CI, MangledName, DemangledName);
-    return;
-  }
-  if (DemangledName == kOCLBuiltinName::ToGlobal ||
-      DemangledName == kOCLBuiltinName::ToLocal ||
-      DemangledName == kOCLBuiltinName::ToPrivate) {
-    visitCallToAddr(&CI, MangledName, DemangledName);
-    return;
-  }
-  if (DemangledName.find(kOCLBuiltinName::VLoadPrefix) == 0 ||
-      DemangledName.find(kOCLBuiltinName::VStorePrefix) == 0) {
-    visitCallVecLoadStore(&CI, MangledName, DemangledName);
-    return;
-  }
-  if (DemangledName == kOCLBuiltinName::IsFinite ||
-      DemangledName == kOCLBuiltinName::IsInf ||
-      DemangledName == kOCLBuiltinName::IsNan ||
-      DemangledName == kOCLBuiltinName::IsNormal ||
-      DemangledName == kOCLBuiltinName::Signbit) {
-    visitCallRelational(&CI, DemangledName);
-    return;
-  }
-  if (DemangledName == kOCLBuiltinName::WorkGroupBarrier ||
-      DemangledName == kOCLBuiltinName::Barrier) {
-    visitCallWorkGroupBarrier(&CI);
-    return;
-  }
-  if (DemangledName == kOCLBuiltinName::GetFence) {
-    visitCallGetFence(&CI, MangledName, DemangledName);
-    return;
-  }
-  if (DemangledName == kOCLBuiltinName::Dot &&
-      !(CI.getOperand(0)->getType()->isVectorTy())) {
-    visitCallDot(&CI);
-    return;
-  }
-  if (DemangledName == kOCLBuiltinName::FMin ||
-      DemangledName == kOCLBuiltinName::FMax ||
-      DemangledName == kOCLBuiltinName::Min ||
-      DemangledName == kOCLBuiltinName::Max ||
-      DemangledName == kOCLBuiltinName::Step ||
-      DemangledName == kOCLBuiltinName::SmoothStep ||
-      DemangledName == kOCLBuiltinName::Clamp ||
-      DemangledName == kOCLBuiltinName::Mix) {
-    visitCallScalToVec(&CI, MangledName, DemangledName);
-    return;
-  }
-  if (DemangledName == kOCLBuiltinName::GetImageChannelDataType) {
-    visitCallGetImageChannel(&CI, MangledName, DemangledName,
-                             OCLImageChannelDataTypeOffset);
-    return;
-  }
-  if (DemangledName == kOCLBuiltinName::GetImageChannelOrder) {
-    visitCallGetImageChannel(&CI, MangledName, DemangledName,
-                             OCLImageChannelOrderOffset);
-    return;
-  }
-  visitCallBuiltinSimple(&CI, MangledName, DemangledName);
-}
-
-void
-OCL20ToSPIRV::visitCallNDRange(CallInst *CI,
-    const std::string &DemangledName) {
-  assert(DemangledName.find(kOCLBuiltinName::NDRangePrefix) == 0);
-  std::string lenStr = DemangledName.substr(8, 1);
-  auto Len = atoi(lenStr.c_str());
-  assert (Len >= 1 && Len <= 3);
-  // SPIR-V ndrange structure requires 3 members in the following order:
-  //   global work offset
-  //   global work size
-  //   local work size
-  // The arguments need to add missing members.
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  mutateCallInstSPIRV(M, CI, [=](CallInst *, std::vector<Value *> &Args){
-    for (size_t I = 1, E = Args.size(); I != E; ++I)
-      Args[I] = getScalarOrArray(Args[I], Len, CI);
-    switch (Args.size()) {
-    case 2: {
-      // Has global work size.
-      auto T = Args[1]->getType();
-      auto C = getScalarOrArrayConstantInt(CI, T, Len, 0);
-      Args.push_back(C);
-      Args.push_back(C);
-    }
-      break;
-    case 3: {
-      // Has global and local work size.
-      auto T = Args[1]->getType();
-      Args.push_back(getScalarOrArrayConstantInt(CI, T, Len, 0));
-    }
-      break;
-    case 4: {
-      // Move offset arg to the end
-      auto OffsetPos = Args.begin() + 1;
-      Value* OffsetVal = *OffsetPos;
-      Args.erase(OffsetPos);
-      Args.push_back(OffsetVal);
-    }
-      break;
-    default:
-      assert(0 && "Invalid number of arguments");
-    }
-    // Translate ndrange_ND into differently named SPIR-V decorated functions because
-    // they have array arugments of different dimension which mangled the same way.
-    return getSPIRVFuncName(OpBuildNDRange, "_" + lenStr + "D");
-  }, &Attrs);
-}
-
-void
-OCL20ToSPIRV::visitCallAsyncWorkGroupCopy(CallInst* CI,
-    const std::string &DemangledName) {
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  mutateCallInstSPIRV(M, CI, [=](CallInst *, std::vector<Value *> &Args){
-    if (DemangledName == OCLUtil::kOCLBuiltinName::AsyncWorkGroupCopy) {
-      Args.insert(Args.begin()+3, addSizet(1));
-    }
-    Args.insert(Args.begin(), addInt32(ScopeWorkgroup));
-    return getSPIRVFuncName(OpGroupAsyncCopy);
-  }, &Attrs);
-}
-
-CallInst *
-OCL20ToSPIRV::visitCallAtomicCmpXchg(CallInst* CI,
-    const std::string& DemangledName) {
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  Value *Expected = nullptr;
-  CallInst *NewCI = nullptr;
-  mutateCallInstOCL(M, CI, [&](CallInst * CI, std::vector<Value *> &Args,
-      Type *&RetTy){
-    Expected = Args[1]; // temporary save second argument.
-    Args[1] = new LoadInst(Args[1], "exp", false, CI);
-    RetTy = Args[2]->getType();
-    assert(Args[0]->getType()->getPointerElementType()->isIntegerTy() &&
-      Args[1]->getType()->isIntegerTy() && Args[2]->getType()->isIntegerTy() &&
-      "In SPIR-V 1.0 arguments of OpAtomicCompareExchange must be "
-      "an integer type scalars");
-    return kOCLBuiltinName::AtomicCmpXchgStrong;
-  },
-  [&](CallInst *NCI)->Instruction * {
-    NewCI = NCI;
-    Instruction* Store = new StoreInst(NCI, Expected, NCI->getNextNode());
-    return new ICmpInst(Store->getNextNode(), CmpInst::ICMP_EQ, NCI,
-                        NCI->getArgOperand(1));
-  },
-  &Attrs);
-  return NewCI;
-}
-
-void
-OCL20ToSPIRV::visitCallAtomicInit(CallInst* CI) {
-  auto ST = new StoreInst(CI->getArgOperand(1), CI->getArgOperand(0), CI);
-  ST->takeName(CI);
-  CI->dropAllReferences();
-  CI->eraseFromParent();
-}
-
-void
-OCL20ToSPIRV::visitCallAllAny(spv::Op OC, CallInst* CI) {
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-
-  auto Args = getArguments(CI);
-  assert(Args.size() == 1);
-
-  auto *ArgTy = Args[0]->getType();
-  auto Zero = Constant::getNullValue(Args[0]->getType());
-
-  auto *Cmp = CmpInst::Create(CmpInst::ICmp, CmpInst::ICMP_SLT, Args[0], Zero,
-                               "cast", CI);
-
-  if (!isa<VectorType>(ArgTy)) {
-    auto *Cast = CastInst::CreateZExtOrBitCast(Cmp, Type::getInt32Ty(*Ctx),
-                                                "", Cmp->getNextNode());
-    CI->replaceAllUsesWith(Cast);
-    CI->eraseFromParent();
-  } else {
-    mutateCallInstSPIRV(
-        M, CI,
-        [&](CallInst *, std::vector<Value *> &Args, Type *&Ret) {
-          Args[0] = Cmp;
-          Ret = Type::getInt1Ty(*Ctx);
-
-          return getSPIRVFuncName(OC);
-        },
-        [&](CallInst *CI) -> Instruction * {
-          return CastInst::CreateZExtOrBitCast(CI, Type::getInt32Ty(*Ctx), "",
-                                               CI->getNextNode());
-        },
-        &Attrs);
-  }
-}
-
-void
-OCL20ToSPIRV::visitCallAtomicWorkItemFence(CallInst* CI) {
-  transMemoryBarrier(CI, getAtomicWorkItemFenceLiterals(CI));
-}
-
-void
-OCL20ToSPIRV::visitCallMemFence(CallInst* CI) {
-  transMemoryBarrier(CI, std::make_tuple(
-      cast<ConstantInt>(CI->getArgOperand(0))->getZExtValue(),
-      OCLMO_relaxed,
-      OCLMS_work_group));
-}
-
-void OCL20ToSPIRV::transMemoryBarrier(CallInst* CI,
-    AtomicWorkItemFenceLiterals Lit) {
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  mutateCallInstSPIRV(M, CI, [=](CallInst *, std::vector<Value *> &Args){
-    Args.resize(2);
-    Args[0] = addInt32(map<Scope>(std::get<2>(Lit)));
-    Args[1] = addInt32(mapOCLMemSemanticToSPIRV(std::get<0>(Lit),
-        std::get<1>(Lit)));
-    return getSPIRVFuncName(OpMemoryBarrier);
-  }, &Attrs);
-}
-
-void
-OCL20ToSPIRV::visitCallAtomicLegacy(CallInst* CI,
-    StringRef MangledName, const std::string& DemangledName) {
-  StringRef Stem = DemangledName;
-  if (Stem.startswith("atom_"))
-    Stem = Stem.drop_front(strlen("atom_"));
-  else if (Stem.startswith("atomic_"))
-    Stem = Stem.drop_front(strlen("atomic_"));
-  else
-    return;
-
-  std::string Sign;
-  std::string Postfix;
-  std::string Prefix;
-  if (Stem == "add" ||
-      Stem == "sub" ||
-      Stem == "and" ||
-      Stem == "or" ||
-      Stem == "xor" ||
-      Stem == "min" ||
-      Stem == "max") {
-    if ((Stem == "min" || Stem == "max") &&
-         isMangledTypeUnsigned(MangledName.back()))
-      Sign = 'u';
-    Prefix = "fetch_";
-    Postfix = "_explicit";
-  } else if (Stem == "xchg") {
-    Stem = "exchange";
-    Postfix = "_explicit";
-  }
-  else if (Stem == "cmpxchg") {
-    Stem = "compare_exchange_strong";
-    Postfix = "_explicit";
-  }
-  else if (Stem == "inc" ||
-           Stem == "dec") {
-    // do nothing
-  } else
-    return;
-
-  OCLBuiltinTransInfo Info;
-  Info.UniqName = "atomic_" + Prefix + Sign + Stem.str() + Postfix;
-  std::vector<int> PostOps;
-  PostOps.push_back(OCLLegacyAtomicMemOrder);
-  if (Stem.startswith("compare_exchange"))
-    PostOps.push_back(OCLLegacyAtomicMemOrder);
-  PostOps.push_back(OCLLegacyAtomicMemScope);
-
-  Info.PostProc = [=](std::vector<Value *> &Ops){
-    for (auto &I:PostOps){
-      Ops.push_back(addInt32(I));
-    }
-  };
-  transAtomicBuiltin(CI, Info);
-}
-
-void
-OCL20ToSPIRV::visitCallAtomicCpp11(CallInst* CI,
-    StringRef MangledName, const std::string& DemangledName) {
-  StringRef Stem = DemangledName;
-  if (Stem.startswith("atomic_"))
-    Stem = Stem.drop_front(strlen("atomic_"));
-  else
-    return;
-
-  std::string NewStem = Stem;
-  std::vector<int> PostOps;
-  if (Stem.startswith("store") ||
-      Stem.startswith("load") ||
-      Stem.startswith("exchange") ||
-      Stem.startswith("compare_exchange") ||
-      Stem.startswith("fetch") ||
-      Stem.startswith("flag")) {
-    if ((Stem.startswith("fetch_min") ||
-        Stem.startswith("fetch_max")) &&
-        containsUnsignedAtomicType(MangledName))
-      NewStem.insert(NewStem.begin() + strlen("fetch_"), 'u');
-
-    if (!Stem.endswith("_explicit")) {
-      NewStem = NewStem + "_explicit";
-      PostOps.push_back(OCLMO_seq_cst);
-      if (Stem.startswith("compare_exchange"))
-        PostOps.push_back(OCLMO_seq_cst);
-      PostOps.push_back(OCLMS_device);
-    } else {
-      auto MaxOps = getOCLCpp11AtomicMaxNumOps(
-          Stem.drop_back(strlen("_explicit")));
-      if (CI->getNumArgOperands() < MaxOps)
-        PostOps.push_back(OCLMS_device);
-    }
-  } else if (Stem == "work_item_fence") {
-    // do nothing
-  } else
-    return;
-
-  OCLBuiltinTransInfo Info;
-  Info.UniqName = std::string("atomic_") + NewStem;
-  Info.PostProc = [=](std::vector<Value *> &Ops){
-    for (auto &I:PostOps){
-      Ops.push_back(addInt32(I));
-    }
-  };
-
-  transAtomicBuiltin(CI, Info);
-}
-
-void
-OCL20ToSPIRV::transAtomicBuiltin(CallInst* CI,
-    OCLBuiltinTransInfo& Info) {
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  mutateCallInstSPIRV(M, CI, [=](CallInst * CI, std::vector<Value *> &Args){
-    Info.PostProc(Args);
-    // Order of args in OCL20:
-    // object, 0-2 other args, 1-2 order, scope
-    const size_t NumOrder = getAtomicBuiltinNumMemoryOrderArgs(Info.UniqName);
-    const size_t ArgsCount = Args.size();
-    const size_t ScopeIdx = ArgsCount - 1;
-    const size_t OrderIdx = ScopeIdx - NumOrder;
-    Args[ScopeIdx] = mapUInt(M, cast<ConstantInt>(Args[ScopeIdx]),
-        [](unsigned I){
-      return map<Scope>(static_cast<OCLScopeKind>(I));
-    });
-    for (size_t I = 0; I < NumOrder; ++I)
-      Args[OrderIdx + I] = mapUInt(M, cast<ConstantInt>(Args[OrderIdx + I]),
-          [](unsigned Ord) {
-      return mapOCLMemSemanticToSPIRV(0, static_cast<OCLMemOrderKind>(Ord));
-    });
-    // Order of args in SPIR-V:
-    // object, scope, 1-2 order, 0-2 other args
-    std::swap(Args[1], Args[ScopeIdx]);
-    if(OrderIdx > 2) {
-      // For atomic_compare_exchange the swap above puts Comparator/Expected
-      // argument just where it should be, so don't move the last argument then.
-      int offset = Info.UniqName.find("atomic_compare_exchange") == 0 ? 1 : 0;
-      std::rotate(Args.begin() + 2, Args.begin() + OrderIdx,
-                  Args.end() - offset);
-    }
-    return getSPIRVFuncName(OCLSPIRVBuiltinMap::map(Info.UniqName));
-  }, &Attrs);
-}
-
-void
-OCL20ToSPIRV::visitCallWorkGroupBarrier(CallInst* CI) {
-  auto Lit = getWorkGroupBarrierLiterals(CI);
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  mutateCallInstSPIRV(M, CI, [=](CallInst *, std::vector<Value *> &Args){
-    Args.resize(3);
-    Args[0] = addInt32(map<Scope>(std::get<2>(Lit)));
-    Args[1] = addInt32(map<Scope>(std::get<1>(Lit)));
-    Args[2] = addInt32(mapOCLMemFenceFlagToSPIRV(std::get<0>(Lit)));
-    return getSPIRVFuncName(OpControlBarrier);
-  }, &Attrs);
-}
-
-void OCL20ToSPIRV::visitCallConvert(CallInst* CI,
-    StringRef MangledName, const std::string& DemangledName) {
-  if (eraseUselessConvert(CI, MangledName, DemangledName))
-    return;
-  Op OC = OpNop;
-  auto TargetTy = CI->getType();
-  auto SrcTy = CI->getArgOperand(0)->getType();
-  if (isa<VectorType>(TargetTy))
-    TargetTy = TargetTy->getVectorElementType();
-  if (isa<VectorType>(SrcTy))
-    SrcTy = SrcTy->getVectorElementType();
-  auto IsTargetInt = isa<IntegerType>(TargetTy);
-
-  std::string TargetTyName = DemangledName.substr(
-      strlen(kOCLBuiltinName::ConvertPrefix));
-  auto FirstUnderscoreLoc = TargetTyName.find('_');
-  if (FirstUnderscoreLoc != std::string::npos)
-    TargetTyName = TargetTyName.substr(0, FirstUnderscoreLoc);
-  TargetTyName = std::string("_R") + TargetTyName;
-
-  std::string Sat = DemangledName.find("_sat") != std::string::npos ?
-      "_sat" : "";
-  auto TargetSigned = DemangledName[8] != 'u';
-  if (isa<IntegerType>(SrcTy)) {
-    bool Signed = isLastFuncParamSigned(MangledName);
-    if (IsTargetInt) {
-      if (!Sat.empty() && TargetSigned != Signed) {
-        OC = Signed ? OpSatConvertSToU : OpSatConvertUToS;
-        Sat = "";
-      } else
-        OC = Signed ? OpSConvert : OpUConvert;
-    } else
-      OC = Signed ? OpConvertSToF : OpConvertUToF;
-  } else {
-    if (IsTargetInt) {
-      OC = TargetSigned ? OpConvertFToS : OpConvertFToU;
-    } else
-      OC = OpFConvert;
-  }
-  auto Loc = DemangledName.find("_rt");
-  std::string Rounding;
-  if (Loc != std::string::npos &&
-      !(isa<IntegerType>(SrcTy) && IsTargetInt)) {
-    Rounding = DemangledName.substr(Loc, 4);
-  }
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  mutateCallInstSPIRV(M, CI, [=](CallInst *, std::vector<Value *> &Args){
-    return getSPIRVFuncName(OC, TargetTyName + Sat + Rounding);
-  }, &Attrs);
-}
-
-void OCL20ToSPIRV::visitCallGroupBuiltin(CallInst* CI,
-    StringRef MangledName, const std::string& OrigDemangledName) {
-  auto F = CI->getCalledFunction();
-  std::vector<int> PreOps;
-  std::string DemangledName = OrigDemangledName;
-
-  if (DemangledName == kOCLBuiltinName::WorkGroupBarrier)
-    return;
-  if (DemangledName == kOCLBuiltinName::WaitGroupEvent) {
-    PreOps.push_back(ScopeWorkgroup);
-  } else if (DemangledName.find(kOCLBuiltinName::WorkGroupPrefix) == 0) {
-    DemangledName.erase(0, strlen(kOCLBuiltinName::WorkPrefix));
-    PreOps.push_back(ScopeWorkgroup);
-  } else if (DemangledName.find(kOCLBuiltinName::SubGroupPrefix) == 0) {
-    DemangledName.erase(0, strlen(kOCLBuiltinName::SubPrefix));
-    PreOps.push_back(ScopeSubgroup);
-  } else
-    return;
-
-  if (DemangledName != kOCLBuiltinName::WaitGroupEvent) {
-    StringRef GroupOp = DemangledName;
-    GroupOp = GroupOp.drop_front(strlen(kSPIRVName::GroupPrefix));
-    SPIRSPIRVGroupOperationMap::foreach_conditional([&](const std::string &S,
-        SPIRVGroupOperationKind G){
-      if (!GroupOp.startswith(S))
-        return true; // continue
-      PreOps.push_back(G);
-      StringRef Op = GroupOp.drop_front(S.size() + 1);
-      assert(!Op.empty() && "Invalid OpenCL group builtin function");
-      char OpTyC = 0;
-      auto NeedSign = Op == "max" || Op == "min";
-      auto OpTy = F->getReturnType();
-      if (OpTy->isFloatingPointTy())
-        OpTyC = 'f';
-      else if (OpTy->isIntegerTy()) {
-        if (!NeedSign)
-          OpTyC = 'i';
-        else {
-          if (isLastFuncParamSigned(F->getName()))
-            OpTyC = 's';
-          else
-            OpTyC = 'u';
-        }
-      } else
-        llvm_unreachable("Invalid OpenCL group builtin argument type");
-
-      DemangledName = std::string(kSPIRVName::GroupPrefix) + OpTyC + Op.str();
-      return false; // break out of loop
-    });
-  }
-
-  bool IsGroupAllAny = (DemangledName.find("_all") != std::string::npos ||
-                        DemangledName.find("_any") != std::string::npos);
-
-  auto Consts = getInt32(M, PreOps);
-  OCLBuiltinTransInfo Info;
-  if (IsGroupAllAny)
-    Info.RetTy = Type::getInt1Ty(*Ctx);
-  Info.UniqName = DemangledName;
-  Info.PostProc = [=](std::vector<Value *> &Ops) {
-    if (IsGroupAllAny) {
-      IRBuilder<> IRB(CI);
-      Ops[0] =
-          IRB.CreateICmpNE(Ops[0], ConstantInt::get(Type::getInt32Ty(*Ctx), 0));
-    }
-    size_t E = Ops.size();
-    if (DemangledName == "group_broadcast" && E > 2) {
-      assert(E == 3 || E == 4);
-      makeVector(CI, Ops, std::make_pair(Ops.begin() + 1, Ops.end()));
-    }
-    Ops.insert(Ops.begin(), Consts.begin(), Consts.end());
-  };
-  transBuiltin(CI, Info);
-}
-
-void
-OCL20ToSPIRV::transBuiltin(CallInst* CI,
-    OCLBuiltinTransInfo& Info) {
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  Op OC = OpNop;
-  unsigned ExtOp = ~0U;
-  if (StringRef(Info.UniqName).startswith(kSPIRVName::Prefix))
-      return;
-  if (OCLSPIRVBuiltinMap::find(Info.UniqName, &OC))
-    Info.UniqName = getSPIRVFuncName(OC);
-  else if ((ExtOp = getExtOp(Info.MangledName, Info.UniqName)) != ~0U)
-    Info.UniqName = getSPIRVExtFuncName(SPIRVEIS_OpenCL, ExtOp);
-  else
-    return;
-  if (!Info.RetTy)
-    mutateCallInstSPIRV(M, CI,
-                        [=](CallInst *, std::vector<Value *> &Args) {
-                          Info.PostProc(Args);
-                          return Info.UniqName + Info.Postfix;
-                        },
-                        &Attrs);
-  else
-    mutateCallInstSPIRV(
-        M, CI,
-        [=](CallInst *, std::vector<Value *> &Args, Type *&RetTy) {
-          Info.PostProc(Args);
-          RetTy = Info.RetTy;
-          return Info.UniqName + Info.Postfix;
-        },
-        [=](CallInst *NewCI) -> Instruction * {
-          if (NewCI->getType()->isIntegerTy() && CI->getType()->isIntegerTy())
-            return CastInst::CreateIntegerCast(NewCI, CI->getType(),
-                                               Info.isRetSigned, "", CI);
-          else
-            return CastInst::CreatePointerBitCastOrAddrSpaceCast(
-                NewCI, CI->getType(), "", CI);
-        },
-        &Attrs);
-}
-
-void
-OCL20ToSPIRV::visitCallPipeBuiltin(CallInst* CI,
-    StringRef MangledName, const std::string& DemangledName) {
-  std::string NewName = DemangledName;
-  // Transform OpenCL read_pipe/write_pipe builtin function names
-  // with reserve_id argument to reserved_read_pipe/reserved_write_pipe.
-  if ((DemangledName.find(kOCLBuiltinName::ReadPipe) == 0 ||
-      DemangledName.find(kOCLBuiltinName::WritePipe) == 0)
-      && CI->getNumArgOperands() > 4)
-    NewName = std::string(kSPIRVName::ReservedPrefix) + DemangledName;
-  OCLBuiltinTransInfo Info;
-  Info.UniqName = NewName;
-  transBuiltin(CI, Info);
-}
-
-void OCL20ToSPIRV::visitCallReadImageMSAA(CallInst *CI, StringRef MangledName,
-                                          const std::string &DemangledName) {
-  assert(MangledName.find("msaa") != StringRef::npos);
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  mutateCallInstSPIRV(
-      M, CI,
-      [=](CallInst *, std::vector<Value *> &Args) {
-        Args.insert(Args.begin() + 2, getInt32(M, ImageOperandsSampleMask));
-        return getSPIRVFuncName(OpImageRead,
-                                std::string(kSPIRVPostfix::ExtDivider) +
-                                    getPostfixForReturnType(CI));
-      },
-      &Attrs);
-}
-
-void OCL20ToSPIRV::visitCallReadImageWithSampler(
-    CallInst *CI, StringRef MangledName, const std::string &DemangledName) {
-  assert (MangledName.find(kMangledName::Sampler) != StringRef::npos);
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  bool isRetScalar = !CI->getType()->isVectorTy();
-  mutateCallInstSPIRV(
-      M, CI,
-      [=](CallInst *, std::vector<Value *> &Args, Type *&Ret) {
-        auto ImageTy = getAnalysis<OCLTypeToSPIRV>().getAdaptedType(Args[0]);
-        if (isOCLImageType(ImageTy))
-          ImageTy = getSPIRVImageTypeFromOCL(M, ImageTy);
-        auto SampledImgTy = getSPIRVTypeByChangeBaseTypeName(
-            M, ImageTy, kSPIRVTypeName::Image, kSPIRVTypeName::SampledImg);
-        Value *SampledImgArgs[] = {Args[0], Args[1]};
-        auto SampledImg = addCallInstSPIRV(
-            M, getSPIRVFuncName(OpSampledImage), SampledImgTy, SampledImgArgs,
-            nullptr, CI, kSPIRVName::TempSampledImage);
-
-        Args[0] = SampledImg;
-        Args.erase(Args.begin() + 1, Args.begin() + 2);
-
-        switch (Args.size()) {
-        case 2: // no lod
-          Args.push_back(getInt32(M, ImageOperandsMask::ImageOperandsLodMask));
-          Args.push_back(getFloat32(M, 0.f));
-          break;
-        case 3: // explicit lod
-          Args.insert(Args.begin() + 2,
-                      getInt32(M, ImageOperandsMask::ImageOperandsLodMask));
-          break;
-        case 4: // gradient
-          Args.insert(Args.begin() + 2,
-                      getInt32(M, ImageOperandsMask::ImageOperandsGradMask));
-          break;
-        default:
-          assert(0 && "read_image* with unhandled number of args!");
-        }
-
-        // SPIR-V intruction always returns 4-element vector
-        if (isRetScalar)
-          Ret = VectorType::get(Ret, 4);
-        return getSPIRVFuncName(OpImageSampleExplicitLod,
-                                std::string(kSPIRVPostfix::ExtDivider) +
-                                    getPostfixForReturnType(Ret));
-      },
-      [&](CallInst *CI) -> Instruction * {
-        if (isRetScalar)
-          return ExtractElementInst::Create(CI, getSizet(M, 0), "",
-                                            CI->getNextNode());
-        return CI;
-      },
-      &Attrs);
-}
-
-void
-OCL20ToSPIRV::visitCallGetImageSize(CallInst* CI,
-    StringRef MangledName, const std::string& DemangledName) {
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  StringRef TyName;
-  auto IsImg = isOCLImageType(CI->getArgOperand(0)->getType(), &TyName);
-  assert(IsImg);
-  SPIRVTypeImageDescriptor Desc = map<SPIRVTypeImageDescriptor>(TyName.str());
-  unsigned Dim = getImageDimension(Desc.Dim) + Desc.Arrayed;
-  assert(Dim > 0 && "Ivalid image dimention.");
-  mutateCallInstSPIRV(M, CI,
-    [&](CallInst *, std::vector<Value *> &Args, Type *&Ret){
-      assert(Args.size() == 1);
-      Ret = CI->getType()->isIntegerTy(64) ? Type::getInt64Ty(*Ctx)
-                                           : Type::getInt32Ty(*Ctx);
-      if (Dim > 1)
-        Ret = VectorType::get(Ret, Dim);
-      if (Desc.Dim == DimBuffer)
-        return getSPIRVFuncName(OpImageQuerySize, CI->getType());
-      else {
-        Args.push_back(getInt32(M, 0));
-        return getSPIRVFuncName(OpImageQuerySizeLod, CI->getType());
-      }
-    },
-    [&](CallInst *NCI)->Instruction * {
-      if (Dim == 1)
-        return NCI;
-      if (DemangledName == kOCLBuiltinName::GetImageDim) {
-        if (Desc.Dim == Dim3D) {
-          auto ZeroVec = ConstantVector::getSplat(3,
-            Constant::getNullValue(NCI->getType()->getVectorElementType()));
-          Constant *Index[] = {getInt32(M, 0), getInt32(M, 1),
-              getInt32(M, 2), getInt32(M, 3)};
-          return new ShuffleVectorInst(NCI, ZeroVec,
-             ConstantVector::get(Index), "", CI);
-
-        } else if (Desc.Dim == Dim2D && Desc.Arrayed) {
-          Constant *Index[] = {getInt32(M, 0), getInt32(M, 1)};
-          Constant *mask = ConstantVector::get(Index);
-          return new ShuffleVectorInst(NCI, UndefValue::get(NCI->getType()),
-                                       mask, NCI->getName(), CI);
-        }
-        return NCI;
-      }
-      unsigned I = StringSwitch<unsigned>(DemangledName)
-          .Case(kOCLBuiltinName::GetImageWidth, 0)
-          .Case(kOCLBuiltinName::GetImageHeight, 1)
-          .Case(kOCLBuiltinName::GetImageDepth, 2)
-          .Case(kOCLBuiltinName::GetImageArraySize, Dim - 1);
-      return ExtractElementInst::Create(NCI, getUInt32(M, I), "",
-          NCI->getNextNode());
-    },
-  &Attrs);
-}
-
-/// Remove trivial conversion functions
-bool
-OCL20ToSPIRV::eraseUselessConvert(CallInst *CI,
-    const std::string &MangledName,
-    const std::string &DemangledName) {
-  auto TargetTy = CI->getType();
-  auto SrcTy = CI->getArgOperand(0)->getType();
-  if (isa<VectorType>(TargetTy))
-    TargetTy = TargetTy->getVectorElementType();
-  if (isa<VectorType>(SrcTy))
-    SrcTy = SrcTy->getVectorElementType();
-  if (TargetTy == SrcTy) {
-    if (isa<IntegerType>(TargetTy) &&
-        DemangledName.find("_sat") != std::string::npos &&
-        isLastFuncParamSigned(MangledName) != (DemangledName[8] != 'u'))
-      return false;
-    CI->getArgOperand(0)->takeName(CI);
-    SPIRVDBG(dbgs() << "[regularizeOCLConvert] " << *CI << " <- " <<
-        *CI->getArgOperand(0) << '\n');
-    CI->replaceAllUsesWith(CI->getArgOperand(0));
-    ValuesToDelete.insert(CI);
-    ValuesToDelete.insert(CI->getCalledFunction());
-    return true;
-  }
-  return false;
-}
-
-void
-OCL20ToSPIRV::visitCallBuiltinSimple(CallInst* CI,
-    StringRef MangledName, const std::string& DemangledName) {
-  OCLBuiltinTransInfo Info;
-  Info.MangledName = MangledName.str();
-  Info.UniqName = DemangledName;
-  transBuiltin(CI, Info);
-}
-
-/// Translates OCL work-item builtin functions to SPIRV builtin variables.
-/// Function like get_global_id(i) -> x = load GlobalInvocationId; extract x, i
-/// Function like get_work_dim() -> load WorkDim
-void OCL20ToSPIRV::transWorkItemBuiltinsToVariables() {
-  DEBUG(dbgs() << "Enter transWorkItemBuiltinsToVariables\n");
-  std::vector<Function *> WorkList;
-  for (auto I = M->begin(), E = M->end(); I != E; ++I) {
-    std::string DemangledName;
-    if (!oclIsBuiltin(I->getName(), &DemangledName))
-      continue;
-    DEBUG(dbgs() << "Function demangled name: " << DemangledName << '\n');
-    std::string BuiltinVarName;
-    SPIRVBuiltinVariableKind BVKind;
-    if (!SPIRSPIRVBuiltinVariableMap::find(DemangledName, &BVKind))
-      continue;
-    BuiltinVarName = std::string(kSPIRVName::Prefix) +
-        SPIRVBuiltInNameMap::map(BVKind);
-    DEBUG(dbgs() << "builtin variable name: " << BuiltinVarName << '\n');
-    bool IsVec = I->getFunctionType()->getNumParams() > 0;
-    Type *GVType = IsVec ? VectorType::get(I->getReturnType(),3) :
-        I->getReturnType();
-    auto BV = new GlobalVariable(*M, GVType,
-        true,
-        GlobalValue::ExternalLinkage,
-        nullptr, BuiltinVarName,
-        0,
-        GlobalVariable::NotThreadLocal,
-        SPIRAS_Constant);
-    std::vector<Instruction *> InstList;
-    for (auto UI = I->user_begin(), UE = I->user_end(); UI != UE; ++UI) {
-      auto CI = dyn_cast<CallInst>(*UI);
-      assert(CI && "invalid instruction");
-      Value * NewValue = new LoadInst(BV, "", CI);
-      DEBUG(dbgs() << "Transform: " << *CI << " => " << *NewValue << '\n');
-      if (IsVec) {
-        NewValue = ExtractElementInst::Create(NewValue,
-          CI->getArgOperand(0),
-          "", CI);
-        DEBUG(dbgs() << *NewValue << '\n');
-      }
-      NewValue->takeName(CI);
-      CI->replaceAllUsesWith(NewValue);
-      InstList.push_back(CI);
-    }
-    for (auto &Inst:InstList) {
-      Inst->dropAllReferences();
-      Inst->removeFromParent();
-    }
-    WorkList.push_back(I);
-  }
-  for (auto &I:WorkList) {
-    I->dropAllReferences();
-    I->removeFromParent();
-  }
-}
-
-void
-OCL20ToSPIRV::visitCallReadWriteImage(CallInst* CI,
-    StringRef MangledName, const std::string& DemangledName) {
-  OCLBuiltinTransInfo Info;
-  if (DemangledName.find(kOCLBuiltinName::ReadImage) == 0)
-    Info.UniqName = kOCLBuiltinName::ReadImage;
-
-  if (DemangledName.find(kOCLBuiltinName::WriteImage) == 0)
-  {
-    Info.UniqName = kOCLBuiltinName::WriteImage;
-    Info.PostProc = [&](std::vector<Value*> &Args) {
-        if (Args.size() == 4) // write with lod
-        {
-            auto Lod = Args[2];
-            Args.erase(Args.begin() + 2);
-            Args.push_back(getInt32(M, ImageOperandsMask::ImageOperandsLodMask));
-            Args.push_back(Lod);
-        }
-    };
-  }
-
-  transBuiltin(CI, Info);
-}
-
-void
-OCL20ToSPIRV::visitCallToAddr(CallInst* CI, StringRef MangledName,
-    const std::string &DemangledName) {
-  auto AddrSpace = static_cast<SPIRAddressSpace>(
-      CI->getType()->getPointerAddressSpace());
-  OCLBuiltinTransInfo Info;
-  Info.UniqName = DemangledName;
-  Info.Postfix = std::string(kSPIRVPostfix::Divider) + "To" +
-      SPIRAddrSpaceCapitalizedNameMap::map(AddrSpace);
-  auto StorageClass = addInt32(SPIRSPIRVAddrSpaceMap::map(AddrSpace));
-  Info.RetTy = getInt8PtrTy(cast<PointerType>(CI->getType()));
-  Info.PostProc = [=](std::vector<Value *> &Ops){
-    auto P = Ops.back();
-    Ops.pop_back();
-    Ops.push_back(castToInt8Ptr(P, CI));
-    Ops.push_back(StorageClass);
-  };
-  transBuiltin(CI, Info);
-}
-
-void OCL20ToSPIRV::visitCallRelational(CallInst *CI,
-                                       const std::string &DemangledName) {
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  Op OC = OpNop;
-  OCLSPIRVBuiltinMap::find(DemangledName, &OC);
-  std::string SPIRVName = getSPIRVFuncName(OC);
-  mutateCallInstSPIRV(
-      M, CI,
-      [=](CallInst *, std::vector<Value *> &Args, Type *&Ret) {
-        Ret = Type::getInt1Ty(*Ctx);
-        if (CI->getOperand(0)->getType()->isVectorTy())
-          Ret = VectorType::get(
-              Type::getInt1Ty(*Ctx),
-              CI->getOperand(0)->getType()->getVectorNumElements());
-        return SPIRVName;
-      },
-      [=](CallInst *NewCI) -> Instruction * {
-        Value *False = nullptr, *True = nullptr;
-        if (NewCI->getType()->isVectorTy()) {
-          Type *IntTy = Type::getInt32Ty(*Ctx);
-          if (cast<VectorType>(NewCI->getOperand(0)->getType())
-                  ->getElementType()
-                  ->isDoubleTy())
-            IntTy = Type::getInt64Ty(*Ctx);
-          if (cast<VectorType>(NewCI->getOperand(0)->getType())
-                  ->getElementType()
-                  ->isHalfTy())
-            IntTy = Type::getInt16Ty(*Ctx);
-          Type *VTy = VectorType::get(IntTy,
-                                      NewCI->getType()->getVectorNumElements());
-          False = Constant::getNullValue(VTy);
-          True = Constant::getAllOnesValue(VTy);
-        } else {
-          False = getInt32(M, 0);
-          True = getInt32(M, 1);
-        }
-        return SelectInst::Create(NewCI, True, False, "", NewCI->getNextNode());
-      },
-      &Attrs);
-}
-
-void
-OCL20ToSPIRV::visitCallVecLoadStore(CallInst* CI,
-    StringRef MangledName, const std::string& OrigDemangledName) {
-  std::vector<int> PreOps;
-  std::string DemangledName = OrigDemangledName;
-  if (DemangledName.find(kOCLBuiltinName::VLoadPrefix) == 0 &&
-      DemangledName != kOCLBuiltinName::VLoadHalf) {
-    SPIRVWord Width = getVecLoadWidth(DemangledName);
-    SPIRVDBG(spvdbgs() << "[visitCallVecLoadStore] DemangledName: " <<
-        DemangledName << " Width: " << Width << '\n');
-    PreOps.push_back(Width);
-  } else if (DemangledName.find(kOCLBuiltinName::RoundingPrefix)
-      != std::string::npos) {
-    auto R = SPIRSPIRVFPRoundingModeMap::map(DemangledName.substr(
-        DemangledName.find(kOCLBuiltinName::RoundingPrefix) + 1, 3));
-    PreOps.push_back(R);
-  }
-
-  if (DemangledName.find(kOCLBuiltinName::VLoadAPrefix) == 0)
-    transVecLoadStoreName(DemangledName, kOCLBuiltinName::VLoadAPrefix, true);
-  else
-    transVecLoadStoreName(DemangledName, kOCLBuiltinName::VLoadPrefix, false);
-
-  if (DemangledName.find(kOCLBuiltinName::VStoreAPrefix) == 0)
-    transVecLoadStoreName(DemangledName, kOCLBuiltinName::VStoreAPrefix, true);
-  else
-    transVecLoadStoreName(DemangledName, kOCLBuiltinName::VStorePrefix, false);
-
-
-  auto Consts = getInt32(M, PreOps);
-  OCLBuiltinTransInfo Info;
-  Info.MangledName = MangledName;
-  Info.UniqName = DemangledName;
-  if (DemangledName.find(kOCLBuiltinName::VLoadPrefix) == 0)
-    Info.Postfix = std::string(kSPIRVPostfix::ExtDivider) +
-      getPostfixForReturnType(CI);
-  Info.PostProc = [=](std::vector<Value *> &Ops){
-    Ops.insert(Ops.end(), Consts.begin(), Consts.end());
-  };
-  transBuiltin(CI, Info);
-}
-
-void OCL20ToSPIRV::visitCallGetFence(CallInst *CI, StringRef MangledName,
-                                     const std::string &DemangledName) {
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  Op OC = OpNop;
-  OCLSPIRVBuiltinMap::find(DemangledName, &OC);
-  std::string SPIRVName = getSPIRVFuncName(OC);
-  mutateCallInstSPIRV(M, CI, [=](CallInst *, std::vector<Value *> &Args,
-                                 Type *&Ret) { return SPIRVName; },
-            [=](CallInst *NewCI) -> Instruction * {
-              return BinaryOperator::CreateLShr(NewCI, getInt32(M, 8), "", CI);
-            },
-            &Attrs);
-}
-
-void OCL20ToSPIRV::visitCallDot(CallInst *CI) {
-  IRBuilder<> Builder(CI);
-  Value *FMulVal = Builder.CreateFMul(CI->getOperand(0), CI->getOperand(1));
-  CI->replaceAllUsesWith(FMulVal);
-  CI->dropAllReferences();
-  CI->removeFromParent();
-}
-
-void OCL20ToSPIRV::visitCallScalToVec(CallInst *CI, StringRef MangledName,
-                                      const std::string &DemangledName) {
-  // Check if all arguments have the same type - it's simple case.
-  auto Uniform = true;
-  auto IsArg0Vector = isa<VectorType>(CI->getOperand(0)->getType());
-  for (unsigned I = 1, E = CI->getNumArgOperands(); Uniform && (I != E); ++I) {
-    Uniform = isa<VectorType>(CI->getOperand(I)->getType()) == IsArg0Vector;
-  }
-  if (Uniform) {
-    visitCallBuiltinSimple(CI, MangledName, DemangledName);
-    return;
-  }
-
-  std::vector<unsigned int> VecPos;
-  std::vector<unsigned int> ScalarPos;
-  if (DemangledName == kOCLBuiltinName::FMin ||
-      DemangledName == kOCLBuiltinName::FMax ||
-      DemangledName == kOCLBuiltinName::Min ||
-      DemangledName == kOCLBuiltinName::Max) {
-    VecPos.push_back(0);
-    ScalarPos.push_back(1);
-  } else if (DemangledName == kOCLBuiltinName::Clamp) {
-    VecPos.push_back(0);
-    ScalarPos.push_back(1);
-    ScalarPos.push_back(2);
-  } else if (DemangledName == kOCLBuiltinName::Mix) {
-    VecPos.push_back(0);
-    VecPos.push_back(1);
-    ScalarPos.push_back(2);
-  } else if (DemangledName == kOCLBuiltinName::Step) {
-    VecPos.push_back(1);
-    ScalarPos.push_back(0);
-  } else if (DemangledName == kOCLBuiltinName::SmoothStep) {
-    VecPos.push_back(2);
-    ScalarPos.push_back(0);
-    ScalarPos.push_back(1);
-  }
-
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  mutateCallInstSPIRV(
-      M, CI,
-      [=](CallInst *, std::vector<Value *> &Args) {
-        Args.resize(VecPos.size() + ScalarPos.size());
-        for (auto I : VecPos) {
-          Args[I] = CI->getOperand(I);
-        }
-        auto VecArgWidth =
-            CI->getOperand(VecPos[0])->getType()->getVectorNumElements();
-        for (auto I : ScalarPos) {
-          Instruction *Inst = InsertElementInst::Create(
-              UndefValue::get(CI->getOperand(VecPos[0])->getType()),
-              CI->getOperand(I), getInt32(M, 0), "", CI);
-          Value *NewVec = new ShuffleVectorInst(
-              Inst, UndefValue::get(CI->getOperand(VecPos[0])->getType()),
-              ConstantVector::getSplat(VecArgWidth, getInt32(M, 0)), "", CI);
-
-          Args[I] = NewVec;
-        }
-        return getSPIRVExtFuncName(SPIRVEIS_OpenCL,
-                                   getExtOp(MangledName, DemangledName));
-      },
-      &Attrs);
-}
-
-void OCL20ToSPIRV::visitCallGetImageChannel(CallInst *CI, StringRef MangledName,
-                                            const std::string &DemangledName,
-                                            unsigned int Offset) {
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  Op OC = OpNop;
-  OCLSPIRVBuiltinMap::find(DemangledName, &OC);
-  std::string SPIRVName = getSPIRVFuncName(OC);
-  mutateCallInstSPIRV(M, CI, [=](CallInst *, std::vector<Value *> &Args,
-                                 Type *&Ret) { return SPIRVName; },
-                      [=](CallInst *NewCI) -> Instruction * {
-                        return BinaryOperator::CreateAdd(
-                            NewCI, getInt32(M, Offset), "", CI);
-                      },
-                      &Attrs);
-}
-}
-
-INITIALIZE_PASS_BEGIN(OCL20ToSPIRV, "cl20tospv", "Transform OCL 2.0 to SPIR-V",
-    false, false)
-INITIALIZE_PASS_DEPENDENCY(OCLTypeToSPIRV)
-INITIALIZE_PASS_END(OCL20ToSPIRV, "cl20tospv", "Transform OCL 2.0 to SPIR-V",
-    false, false)
-
-ModulePass *llvm::createOCL20ToSPIRV() {
-  return new OCL20ToSPIRV();
-}
+//===- OCL20ToSPIRV.cpp - Transform OCL20 to SPIR-V builtins -----*- C++ -*-===//
+//
+//                     The LLVM/SPIRV Translator
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+// Copyright (c) 2014 Advanced Micro Devices, Inc. All rights reserved.
+//
+// Permission is hereby granted, free of charge, to any person obtaining a
+// copy of this software and associated documentation files (the "Software"),
+// to deal with the Software without restriction, including without limitation
+// the rights to use, copy, modify, merge, publish, distribute, sublicense,
+// and/or sell copies of the Software, and to permit persons to whom the
+// Software is furnished to do so, subject to the following conditions:
+//
+// Redistributions of source code must retain the above copyright notice,
+// this list of conditions and the following disclaimers.
+// Redistributions in binary form must reproduce the above copyright notice,
+// this list of conditions and the following disclaimers in the documentation
+// and/or other materials provided with the distribution.
+// Neither the names of Advanced Micro Devices, Inc., nor the names of its
+// contributors may be used to endorse or promote products derived from this
+// Software without specific prior written permission.
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+// CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+// THE SOFTWARE.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements translation of OCL20 builtin functions.
+//
+//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "cl20tospv"
+
+#include "SPIRVInternal.h"
+#include "OCLUtil.h"
+#include "OCLTypeToSPIRV.h"
+
+#include "llvm/ADT/StringSwitch.h"
+#include "llvm/IR/InstVisitor.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/Instruction.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/Verifier.h"
+#include "llvm/Pass.h"
+#include "llvm/PassSupport.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+
+#include <set>
+
+using namespace llvm;
+using namespace SPIRV;
+using namespace OCLUtil;
+
+namespace SPIRV {
+static size_t
+getOCLCpp11AtomicMaxNumOps(StringRef Name) {
+  return StringSwitch<size_t>(Name)
+      .Cases("load", "flag_test_and_set", "flag_clear", 3)
+      .Cases("store", "exchange",  4)
+      .StartsWith("compare_exchange", 6)
+      .StartsWith("fetch", 4)
+      .Default(0);
+}
+
+class OCL20ToSPIRV: public ModulePass,
+  public InstVisitor<OCL20ToSPIRV> {
+public:
+  OCL20ToSPIRV():ModulePass(ID), M(nullptr), Ctx(nullptr), CLVer(0) {
+    initializeOCL20ToSPIRVPass(*PassRegistry::getPassRegistry());
+  }
+  virtual bool runOnModule(Module &M);
+
+  void getAnalysisUsage(AnalysisUsage &AU) const {
+    AU.addRequired<OCLTypeToSPIRV>();
+  }
+
+  virtual void visitCallInst(CallInst &CI);
+
+  /// Transform barrier/work_group_barrier to __spirv_ControlBarrier.
+  /// barrier(flag) =>
+  ///   __spirv_ControlBarrier(workgroup, workgroup, map(flag))
+  /// workgroup_barrier(scope, flag) =>
+  ///   __spirv_ControlBarrier(workgroup, map(scope), map(flag))
+  void visitCallWorkGroupBarrier(CallInst *CI);
+
+  /// Erase useless convert functions.
+  /// \return true if the call instruction is erased.
+  bool eraseUselessConvert(CallInst *Call, const std::string &MangledName,
+      const std::string &DeMangledName);
+
+  /// Transform convert_ to
+  ///   __spirv_{CastOpName}_R{TargeTyName}{_sat}{_rt[p|n|z|e]}
+  void visitCallConvert(CallInst *CI, StringRef MangledName,
+    const std::string &DemangledName);
+
+  /// Transform async_work_group{_strided}_copy.
+  /// async_work_group_copy(dst, src, n, event)
+  ///   => async_work_group_strided_copy(dst, src, n, 1, event)
+  /// async_work_group_strided_copy(dst, src, n, stride, event)
+  ///   => __spirv_AsyncGroupCopy(ScopeWorkGroup, dst, src, n, stride, event)
+  void visitCallAsyncWorkGroupCopy(CallInst *CI,
+      const std::string &DemangledName);
+
+  /// Transform OCL builtin function to SPIR-V builtin function.
+  void transBuiltin(CallInst *CI, OCLBuiltinTransInfo &Info);
+
+  /// Transform OCL work item builtin functions to SPIR-V builtin variables.
+  void transWorkItemBuiltinsToVariables();
+
+  /// Transform atomic_work_item_fence/mem_fence to __spirv_MemoryBarrier.
+  /// func(flag, order, scope) =>
+  ///   __spirv_MemoryBarrier(map(scope), map(flag)|map(order))
+  void transMemoryBarrier(CallInst *CI, AtomicWorkItemFenceLiterals);
+
+  /// Transform all to __spirv_Op(All|Any).  Note that the types mismatch so
+  // some extra code is emitted to convert between the two.
+  void visitCallAllAny(spv::Op OC, CallInst *CI);
+
+  /// Transform atomic_* to __spirv_Atomic*.
+  /// atomic_x(ptr_arg, args, order, scope) =>
+  ///   __spirv_AtomicY(ptr_arg, map(order), map(scope), args)
+  void transAtomicBuiltin(CallInst *CI, OCLBuiltinTransInfo &Info);
+
+  /// Transform atomic_work_item_fence to __spirv_MemoryBarrier.
+  /// atomic_work_item_fence(flag, order, scope) =>
+  ///   __spirv_MemoryBarrier(map(scope), map(flag)|map(order))
+  void visitCallAtomicWorkItemFence(CallInst *CI);
+
+  /// Transform atomic_compare_exchange call.
+  /// In atomic_compare_exchange, the expected value parameter is a pointer.
+  /// However in SPIR-V it is a value. The transformation adds a load 
+  /// instruction, result of which is passed to atomic_compare_exchange as
+  /// argument.
+  /// The transformation adds a store instruction after the call, to update the
+  /// value in expected with the value pointed to by object. Though, it is not
+  /// necessary in case they are equal, this approach makes result code simpler.
+  /// Also ICmp instruction is added, because the call must return result of
+  /// comparison.
+  /// \returns the call instruction of atomic_compare_exchange_strong.
+  CallInst *visitCallAtomicCmpXchg(CallInst *CI,
+      const std::string &DemangledName);
+
+  /// Transform atomic_init.
+  /// atomic_init(p, x) => store p, x
+  void visitCallAtomicInit(CallInst *CI);
+
+  /// Transform legacy OCL 1.x atomic builtins to SPIR-V builtins for extensions
+  ///   cl_khr_int64_base_atomics
+  ///   cl_khr_int64_extended_atomics
+  /// Do nothing if the called function is not a legacy atomic builtin.
+  void visitCallAtomicLegacy(CallInst *CI, StringRef MangledName,
+    const std::string &DemangledName);
+
+  /// Transform OCL 2.0 C++11 atomic builtins to SPIR-V builtins.
+  /// Do nothing if the called function is not a C++11 atomic builtin.
+  void visitCallAtomicCpp11(CallInst *CI, StringRef MangledName,
+    const std::string &DemangledName);
+
+  /// Transform OCL builtin function to SPIR-V builtin function.
+  /// Assuming there is a simple name mapping without argument changes.
+  /// Should be called at last.
+  void visitCallBuiltinSimple(CallInst *CI, StringRef MangledName,
+    const std::string &DemangledName);
+
+  /// Transform get_image_{width|height|depth|dim}.
+  /// get_image_xxx(...) =>
+  ///   dimension = __spirv_ImageQuerySizeLod_R{ReturnType}(...);
+  ///   return dimension.{x|y|z};
+  void visitCallGetImageSize(CallInst *CI, StringRef MangledName,
+    const std::string &DemangledName);
+
+  /// Transform {work|sub}_group_x =>
+  ///   __spirv_{OpName}
+  ///
+  /// Special handling of work_group_broadcast.
+  ///   work_group_broadcast(a, x, y, z)
+  ///     =>
+  ///   __spirv_GroupBroadcast(a, vec3(x, y, z))
+
+  void visitCallGroupBuiltin(CallInst *CI, StringRef MangledName,
+    const std::string &DemangledName);
+
+  /// Transform mem_fence to __spirv_MemoryBarrier.
+  /// mem_fence(flag) => __spirv_MemoryBarrier(Workgroup, map(flag))
+  void visitCallMemFence(CallInst *CI);
+
+  void visitCallNDRange(CallInst *CI, const std::string &DemangledName);
+
+  /// Transform OCL pipe builtin function to SPIR-V pipe builtin function.
+  void visitCallPipeBuiltin(CallInst *CI, StringRef MangledName,
+    const std::string &DemangledName);
+
+  /// Transform read_image with sampler arguments.
+  /// read_image(image, sampler, ...) =>
+  ///   sampled_image = __spirv_SampledImage(image, sampler);
+  ///   return __spirv_ImageSampleExplicitLod_R{ReturnType}(sampled_image, ...);
+  void visitCallReadImageWithSampler(CallInst *CI, StringRef MangledName,
+      const std::string &DemangledName);
+
+  /// Transform read_image with msaa image arguments.
+  /// Sample argument must be acoded as Image Operand.
+  void visitCallReadImageMSAA(CallInst *CI, StringRef MangledName,
+                              const std::string &DemangledName);
+
+  /// Transform {read|write}_image without sampler arguments.
+  void visitCallReadWriteImage(CallInst *CI, StringRef MangledName,
+      const std::string &DemangledName);
+
+  /// Transform to_{global|local|private}.
+  ///
+  /// T* a = ...;
+  /// addr T* b = to_addr(a);
+  ///   =>
+  /// i8* x = cast<i8*>(a);
+  /// addr i8* y = __spirv_GenericCastToPtr_ToAddr(x);
+  /// addr T* b = cast<addr T*>(y);
+  void visitCallToAddr(CallInst *CI, StringRef MangledName,
+      const std::string &DemangledName);
+
+  /// Transform return type of relatinal built-in functions like isnan, isfinite
+  /// to boolean values.
+  void visitCallRelational(CallInst *CI, const std::string &DemangledName);
+
+  /// Transform vector load/store functions to SPIR-V extended builtin
+  ///   functions
+  /// {vload|vstore{a}}{_half}{n}{_rte|_rtz|_rtp|_rtn} =>
+  ///   __spirv_ocl_{ExtendedInstructionOpCodeName}__R{ReturnType}
+  void visitCallVecLoadStore(CallInst *CI, StringRef MangledName,
+      const std::string &DemangledName);
+
+  /// Transforms get_mem_fence built-in to SPIR-V function and aligns result values with SPIR 1.2.
+  /// get_mem_fence(ptr) => __spirv_GenericPtrMemSemantics
+  /// GenericPtrMemSemantics valid values are 0x100, 0x200 and 0x300, where is
+  /// SPIR 1.2 defines them as 0x1, 0x2 and 0x3, so this function adjusts
+  /// GenericPtrMemSemantics results to SPIR 1.2 values.
+  void visitCallGetFence(CallInst *CI, StringRef MangledName, const std::string& DemangledName);
+
+  /// Transforms OpDot instructions with a scalar type to a fmul instruction
+  void visitCallDot(CallInst *CI);
+
+  /// Fixes for built-in functions with vector+scalar arguments that are
+  /// translated to the SPIR-V instructions where all arguments must have the
+  /// same type.
+  void visitCallScalToVec(CallInst *CI, StringRef MangledName,
+                          const std::string &DemangledName);
+
+  /// Transform get_image_channel_{order|data_type} built-in functions to
+  ///   __spirv_ocl_{ImageQueryOrder|ImageQueryFormat}
+  void visitCallGetImageChannel(CallInst *CI, StringRef MangledName,
+                                const std::string &DemangledName,
+                                unsigned int Offset);
+
+  void visitDbgInfoIntrinsic(DbgInfoIntrinsic &I){
+    I.dropAllReferences();
+    I.eraseFromParent();
+  }
+  static char ID;
+private:
+  Module *M;
+  LLVMContext *Ctx;
+  unsigned CLVer;                   /// OpenCL version as major*10+minor
+  std::set<Value *> ValuesToDelete;
+
+  ConstantInt *addInt32(int I) {
+    return getInt32(M, I);
+  }
+  ConstantInt *addSizet(uint64_t I) {
+    return getSizet(M, I);
+  }
+
+  /// Get vector width from OpenCL vload* function name.
+  SPIRVWord getVecLoadWidth(const std::string& DemangledName) {
+    SPIRVWord Width = 0;
+    if (DemangledName == "vloada_half")
+      Width = 1;
+    else {
+      unsigned Loc = 5;
+      if (DemangledName.find("vload_half") == 0)
+        Loc = 10;
+      else if (DemangledName.find("vloada_half") == 0)
+        Loc = 11;
+
+      std::stringstream SS(DemangledName.substr(Loc));
+      SS >> Width;
+    }
+    return Width;
+  }
+
+  /// Transform OpenCL vload/vstore function name.
+  void transVecLoadStoreName(std::string& DemangledName,
+      const std::string &Stem, bool AlwaysN) {
+    auto HalfStem = Stem + "_half";
+    auto HalfStemR = HalfStem + "_r";
+    if (!AlwaysN && DemangledName == HalfStem)
+      return;
+    if (!AlwaysN && DemangledName.find(HalfStemR) == 0) {
+      DemangledName = HalfStemR;
+      return;
+    }
+    if (DemangledName.find(HalfStem) == 0) {
+      auto OldName = DemangledName;
+      DemangledName = HalfStem + "n";
+      if (OldName.find("_r") != std::string::npos)
+        DemangledName += "_r";
+      return;
+    }
+    if (DemangledName.find(Stem) == 0) {
+      DemangledName = Stem + "n";
+      return;
+    }
+  }
+
+};
+
+char OCL20ToSPIRV::ID = 0;
+
+bool
+OCL20ToSPIRV::runOnModule(Module& Module) {
+  M = &Module;
+  Ctx = &M->getContext();
+  auto Src = getSPIRVSource(&Module);
+  if (std::get<0>(Src) != spv::SourceLanguageOpenCL_C)
+    return false;
+
+  CLVer = std::get<1>(Src);
+  if (CLVer > kOCLVer::CL20)
+    return false;
+
+  DEBUG(dbgs() << "Enter OCL20ToSPIRV:\n");
+
+  transWorkItemBuiltinsToVariables();
+
+  visit(*M);
+
+  for (auto &I:ValuesToDelete)
+    if (auto Inst = dyn_cast<Instruction>(I))
+      Inst->eraseFromParent();
+  for (auto &I:ValuesToDelete)
+    if (auto GV = dyn_cast<GlobalValue>(I))
+      GV->eraseFromParent();
+
+  DEBUG(dbgs() << "After OCL20ToSPIRV:\n" << *M);
+
+  std::string Err;
+  raw_string_ostream ErrorOS(Err);
+  if (verifyModule(*M, &ErrorOS)){
+    DEBUG(errs() << "Fails to verify module: " << ErrorOS.str());
+  }
+  return true;
+}
+
+// The order of handling OCL builtin functions is important.
+// Workgroup functions need to be handled before pipe functions since
+// there are functions fall into both categories.
+void
+OCL20ToSPIRV::visitCallInst(CallInst& CI) {
+  DEBUG(dbgs() << "[visistCallInst] " << CI << '\n');
+  auto F = CI.getCalledFunction();
+  if (!F)
+    return;
+
+  auto MangledName = F->getName();
+  std::string DemangledName;
+  if (!oclIsBuiltin(MangledName, &DemangledName))
+    return;
+  DEBUG(dbgs() << "DemangledName: " << DemangledName << '\n');
+  if (DemangledName.find(kOCLBuiltinName::NDRangePrefix) == 0) {
+    visitCallNDRange(&CI, DemangledName);
+    return;
+  }
+  if (DemangledName == kOCLBuiltinName::All) {
+      visitCallAllAny(OpAll, &CI);
+      return;
+  }
+  if (DemangledName == kOCLBuiltinName::Any) {
+      visitCallAllAny(OpAny, &CI);
+      return;
+  }
+  if (DemangledName.find(kOCLBuiltinName::AsyncWorkGroupCopy) == 0 ||
+      DemangledName.find(kOCLBuiltinName::AsyncWorkGroupStridedCopy) == 0) {
+    visitCallAsyncWorkGroupCopy(&CI, DemangledName);
+    return;
+  }
+  if (DemangledName.find(kOCLBuiltinName::AtomicPrefix) == 0 ||
+      DemangledName.find(kOCLBuiltinName::AtomPrefix) == 0) {
+    auto PCI = &CI;
+    if (DemangledName == kOCLBuiltinName::AtomicInit) {
+      visitCallAtomicInit(PCI);
+      return;
+    }
+    if (DemangledName == kOCLBuiltinName::AtomicWorkItemFence) {
+      visitCallAtomicWorkItemFence(PCI);
+      return;
+    }
+    if (DemangledName == kOCLBuiltinName::AtomicCmpXchgWeak ||
+        DemangledName == kOCLBuiltinName::AtomicCmpXchgStrong ||
+        DemangledName == kOCLBuiltinName::AtomicCmpXchgWeakExplicit ||
+        DemangledName == kOCLBuiltinName::AtomicCmpXchgStrongExplicit) {
+      assert(CLVer == kOCLVer::CL20 && "Wrong version of OpenCL");
+      PCI = visitCallAtomicCmpXchg(PCI, DemangledName);
+    }
+    visitCallAtomicLegacy(PCI, MangledName, DemangledName);
+    visitCallAtomicCpp11(PCI, MangledName, DemangledName);
+    return;
+  }
+  if (DemangledName.find(kOCLBuiltinName::ConvertPrefix) == 0) {
+    visitCallConvert(&CI, MangledName, DemangledName);
+    return;
+  }
+  if (DemangledName == kOCLBuiltinName::GetImageWidth ||
+      DemangledName == kOCLBuiltinName::GetImageHeight ||
+      DemangledName == kOCLBuiltinName::GetImageDepth ||
+      DemangledName == kOCLBuiltinName::GetImageDim   ||
+      DemangledName == kOCLBuiltinName::GetImageArraySize) {
+    visitCallGetImageSize(&CI, MangledName, DemangledName);
+    return;
+  }
+  if ((DemangledName.find(kOCLBuiltinName::WorkGroupPrefix) == 0 &&
+      DemangledName != kOCLBuiltinName::WorkGroupBarrier) ||
+      DemangledName == kOCLBuiltinName::WaitGroupEvent ||
+      DemangledName.find(kOCLBuiltinName::SubGroupPrefix) == 0) {
+    visitCallGroupBuiltin(&CI, MangledName, DemangledName);
+    return;
+  }
+  if (DemangledName.find(kOCLBuiltinName::Pipe) != std::string::npos) {
+    visitCallPipeBuiltin(&CI, MangledName, DemangledName);
+    return;
+  }
+  if (DemangledName == kOCLBuiltinName::MemFence) {
+    visitCallMemFence(&CI);
+    return;
+  }
+  if (DemangledName.find(kOCLBuiltinName::ReadImage) == 0) {
+    if (MangledName.find(kMangledName::Sampler) != StringRef::npos) {
+      visitCallReadImageWithSampler(&CI, MangledName, DemangledName);
+      return;
+    }
+    if (MangledName.find("msaa") != StringRef::npos) {
+      visitCallReadImageMSAA(&CI, MangledName, DemangledName);
+      return;
+    }
+  }
+  if (DemangledName.find(kOCLBuiltinName::ReadImage) == 0 ||
+      DemangledName.find(kOCLBuiltinName::WriteImage) == 0) {
+    visitCallReadWriteImage(&CI, MangledName, DemangledName);
+    return;
+  }
+  if (DemangledName == kOCLBuiltinName::ToGlobal ||
+      DemangledName == kOCLBuiltinName::ToLocal ||
+      DemangledName == kOCLBuiltinName::ToPrivate) {
+    visitCallToAddr(&CI, MangledName, DemangledName);
+    return;
+  }
+  if (DemangledName.find(kOCLBuiltinName::VLoadPrefix) == 0 ||
+      DemangledName.find(kOCLBuiltinName::VStorePrefix) == 0) {
+    visitCallVecLoadStore(&CI, MangledName, DemangledName);
+    return;
+  }
+  if (DemangledName == kOCLBuiltinName::IsFinite ||
+      DemangledName == kOCLBuiltinName::IsInf ||
+      DemangledName == kOCLBuiltinName::IsNan ||
+      DemangledName == kOCLBuiltinName::IsNormal ||
+      DemangledName == kOCLBuiltinName::Signbit) {
+    visitCallRelational(&CI, DemangledName);
+    return;
+  }
+  if (DemangledName == kOCLBuiltinName::WorkGroupBarrier ||
+      DemangledName == kOCLBuiltinName::Barrier) {
+    visitCallWorkGroupBarrier(&CI);
+    return;
+  }
+  if (DemangledName == kOCLBuiltinName::GetFence) {
+    visitCallGetFence(&CI, MangledName, DemangledName);
+    return;
+  }
+  if (DemangledName == kOCLBuiltinName::Dot &&
+      !(CI.getOperand(0)->getType()->isVectorTy())) {
+    visitCallDot(&CI);
+    return;
+  }
+  if (DemangledName == kOCLBuiltinName::FMin ||
+      DemangledName == kOCLBuiltinName::FMax ||
+      DemangledName == kOCLBuiltinName::Min ||
+      DemangledName == kOCLBuiltinName::Max ||
+      DemangledName == kOCLBuiltinName::Step ||
+      DemangledName == kOCLBuiltinName::SmoothStep ||
+      DemangledName == kOCLBuiltinName::Clamp ||
+      DemangledName == kOCLBuiltinName::Mix) {
+    visitCallScalToVec(&CI, MangledName, DemangledName);
+    return;
+  }
+  if (DemangledName == kOCLBuiltinName::GetImageChannelDataType) {
+    visitCallGetImageChannel(&CI, MangledName, DemangledName,
+                             OCLImageChannelDataTypeOffset);
+    return;
+  }
+  if (DemangledName == kOCLBuiltinName::GetImageChannelOrder) {
+    visitCallGetImageChannel(&CI, MangledName, DemangledName,
+                             OCLImageChannelOrderOffset);
+    return;
+  }
+  visitCallBuiltinSimple(&CI, MangledName, DemangledName);
+}
+
+void
+OCL20ToSPIRV::visitCallNDRange(CallInst *CI,
+    const std::string &DemangledName) {
+  assert(DemangledName.find(kOCLBuiltinName::NDRangePrefix) == 0);
+  std::string lenStr = DemangledName.substr(8, 1);
+  auto Len = atoi(lenStr.c_str());
+  assert (Len >= 1 && Len <= 3);
+  // SPIR-V ndrange structure requires 3 members in the following order:
+  //   global work offset
+  //   global work size
+  //   local work size
+  // The arguments need to add missing members.
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  mutateCallInstSPIRV(M, CI, [=](CallInst *, std::vector<Value *> &Args){
+    for (size_t I = 1, E = Args.size(); I != E; ++I)
+      Args[I] = getScalarOrArray(Args[I], Len, CI);
+    switch (Args.size()) {
+    case 2: {
+      // Has global work size.
+      auto T = Args[1]->getType();
+      auto C = getScalarOrArrayConstantInt(CI, T, Len, 0);
+      Args.push_back(C);
+      Args.push_back(C);
+    }
+      break;
+    case 3: {
+      // Has global and local work size.
+      auto T = Args[1]->getType();
+      Args.push_back(getScalarOrArrayConstantInt(CI, T, Len, 0));
+    }
+      break;
+    case 4: {
+      // Move offset arg to the end
+      auto OffsetPos = Args.begin() + 1;
+      Value* OffsetVal = *OffsetPos;
+      Args.erase(OffsetPos);
+      Args.push_back(OffsetVal);
+    }
+      break;
+    default:
+      assert(0 && "Invalid number of arguments");
+    }
+    // Translate ndrange_ND into differently named SPIR-V decorated functions because
+    // they have array arugments of different dimension which mangled the same way.
+    return getSPIRVFuncName(OpBuildNDRange, "_" + lenStr + "D");
+  }, &Attrs);
+}
+
+void
+OCL20ToSPIRV::visitCallAsyncWorkGroupCopy(CallInst* CI,
+    const std::string &DemangledName) {
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  mutateCallInstSPIRV(M, CI, [=](CallInst *, std::vector<Value *> &Args){
+    if (DemangledName == OCLUtil::kOCLBuiltinName::AsyncWorkGroupCopy) {
+      Args.insert(Args.begin()+3, addSizet(1));
+    }
+    Args.insert(Args.begin(), addInt32(ScopeWorkgroup));
+    return getSPIRVFuncName(OpGroupAsyncCopy);
+  }, &Attrs);
+}
+
+CallInst *
+OCL20ToSPIRV::visitCallAtomicCmpXchg(CallInst* CI,
+    const std::string& DemangledName) {
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  Value *Expected = nullptr;
+  CallInst *NewCI = nullptr;
+  mutateCallInstOCL(M, CI, [&](CallInst * CI, std::vector<Value *> &Args,
+      Type *&RetTy){
+    Expected = Args[1]; // temporary save second argument.
+    Args[1] = new LoadInst(Args[1], "exp", false, CI);
+    RetTy = Args[2]->getType();
+    assert(Args[0]->getType()->getPointerElementType()->isIntegerTy() &&
+      Args[1]->getType()->isIntegerTy() && Args[2]->getType()->isIntegerTy() &&
+      "In SPIR-V 1.0 arguments of OpAtomicCompareExchange must be "
+      "an integer type scalars");
+    return kOCLBuiltinName::AtomicCmpXchgStrong;
+  },
+  [&](CallInst *NCI)->Instruction * {
+    NewCI = NCI;
+    Instruction* Store = new StoreInst(NCI, Expected, NCI->getNextNode());
+    return new ICmpInst(Store->getNextNode(), CmpInst::ICMP_EQ, NCI,
+                        NCI->getArgOperand(1));
+  },
+  &Attrs);
+  return NewCI;
+}
+
+void
+OCL20ToSPIRV::visitCallAtomicInit(CallInst* CI) {
+  auto ST = new StoreInst(CI->getArgOperand(1), CI->getArgOperand(0), CI);
+  ST->takeName(CI);
+  CI->dropAllReferences();
+  CI->eraseFromParent();
+}
+
+void
+OCL20ToSPIRV::visitCallAllAny(spv::Op OC, CallInst* CI) {
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+
+  auto Args = getArguments(CI);
+  assert(Args.size() == 1);
+
+  auto *ArgTy = Args[0]->getType();
+  auto Zero = Constant::getNullValue(Args[0]->getType());
+
+  auto *Cmp = CmpInst::Create(CmpInst::ICmp, CmpInst::ICMP_SLT, Args[0], Zero,
+                               "cast", CI);
+
+  if (!isa<VectorType>(ArgTy)) {
+    auto *Cast = CastInst::CreateZExtOrBitCast(Cmp, Type::getInt32Ty(*Ctx),
+                                                "", Cmp->getNextNode());
+    CI->replaceAllUsesWith(Cast);
+    CI->eraseFromParent();
+  } else {
+    mutateCallInstSPIRV(
+        M, CI,
+        [&](CallInst *, std::vector<Value *> &Args, Type *&Ret) {
+          Args[0] = Cmp;
+          Ret = Type::getInt1Ty(*Ctx);
+
+          return getSPIRVFuncName(OC);
+        },
+        [&](CallInst *CI) -> Instruction * {
+          return CastInst::CreateZExtOrBitCast(CI, Type::getInt32Ty(*Ctx), "",
+                                               CI->getNextNode());
+        },
+        &Attrs);
+  }
+}
+
+void
+OCL20ToSPIRV::visitCallAtomicWorkItemFence(CallInst* CI) {
+  transMemoryBarrier(CI, getAtomicWorkItemFenceLiterals(CI));
+}
+
+void
+OCL20ToSPIRV::visitCallMemFence(CallInst* CI) {
+  transMemoryBarrier(CI, std::make_tuple(
+      cast<ConstantInt>(CI->getArgOperand(0))->getZExtValue(),
+      OCLMO_relaxed,
+      OCLMS_work_group));
+}
+
+void OCL20ToSPIRV::transMemoryBarrier(CallInst* CI,
+    AtomicWorkItemFenceLiterals Lit) {
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  mutateCallInstSPIRV(M, CI, [=](CallInst *, std::vector<Value *> &Args){
+    Args.resize(2);
+    Args[0] = addInt32(map<Scope>(std::get<2>(Lit)));
+    Args[1] = addInt32(mapOCLMemSemanticToSPIRV(std::get<0>(Lit),
+        std::get<1>(Lit)));
+    return getSPIRVFuncName(OpMemoryBarrier);
+  }, &Attrs);
+}
+
+void
+OCL20ToSPIRV::visitCallAtomicLegacy(CallInst* CI,
+    StringRef MangledName, const std::string& DemangledName) {
+  StringRef Stem = DemangledName;
+  if (Stem.startswith("atom_"))
+    Stem = Stem.drop_front(strlen("atom_"));
+  else if (Stem.startswith("atomic_"))
+    Stem = Stem.drop_front(strlen("atomic_"));
+  else
+    return;
+
+  std::string Sign;
+  std::string Postfix;
+  std::string Prefix;
+  if (Stem == "add" ||
+      Stem == "sub" ||
+      Stem == "and" ||
+      Stem == "or" ||
+      Stem == "xor" ||
+      Stem == "min" ||
+      Stem == "max") {
+    if ((Stem == "min" || Stem == "max") &&
+         isMangledTypeUnsigned(MangledName.back()))
+      Sign = 'u';
+    Prefix = "fetch_";
+    Postfix = "_explicit";
+  } else if (Stem == "xchg") {
+    Stem = "exchange";
+    Postfix = "_explicit";
+  }
+  else if (Stem == "cmpxchg") {
+    Stem = "compare_exchange_strong";
+    Postfix = "_explicit";
+  }
+  else if (Stem == "inc" ||
+           Stem == "dec") {
+    // do nothing
+  } else
+    return;
+
+  OCLBuiltinTransInfo Info;
+  Info.UniqName = "atomic_" + Prefix + Sign + Stem.str() + Postfix;
+  std::vector<int> PostOps;
+  PostOps.push_back(OCLLegacyAtomicMemOrder);
+  if (Stem.startswith("compare_exchange"))
+    PostOps.push_back(OCLLegacyAtomicMemOrder);
+  PostOps.push_back(OCLLegacyAtomicMemScope);
+
+  Info.PostProc = [=](std::vector<Value *> &Ops){
+    for (auto &I:PostOps){
+      Ops.push_back(addInt32(I));
+    }
+  };
+  transAtomicBuiltin(CI, Info);
+}
+
+void
+OCL20ToSPIRV::visitCallAtomicCpp11(CallInst* CI,
+    StringRef MangledName, const std::string& DemangledName) {
+  StringRef Stem = DemangledName;
+  if (Stem.startswith("atomic_"))
+    Stem = Stem.drop_front(strlen("atomic_"));
+  else
+    return;
+
+  std::string NewStem = Stem;
+  std::vector<int> PostOps;
+  if (Stem.startswith("store") ||
+      Stem.startswith("load") ||
+      Stem.startswith("exchange") ||
+      Stem.startswith("compare_exchange") ||
+      Stem.startswith("fetch") ||
+      Stem.startswith("flag")) {
+    if ((Stem.startswith("fetch_min") ||
+        Stem.startswith("fetch_max")) &&
+        containsUnsignedAtomicType(MangledName))
+      NewStem.insert(NewStem.begin() + strlen("fetch_"), 'u');
+
+    if (!Stem.endswith("_explicit")) {
+      NewStem = NewStem + "_explicit";
+      PostOps.push_back(OCLMO_seq_cst);
+      if (Stem.startswith("compare_exchange"))
+        PostOps.push_back(OCLMO_seq_cst);
+      PostOps.push_back(OCLMS_device);
+    } else {
+      auto MaxOps = getOCLCpp11AtomicMaxNumOps(
+          Stem.drop_back(strlen("_explicit")));
+      if (CI->getNumArgOperands() < MaxOps)
+        PostOps.push_back(OCLMS_device);
+    }
+  } else if (Stem == "work_item_fence") {
+    // do nothing
+  } else
+    return;
+
+  OCLBuiltinTransInfo Info;
+  Info.UniqName = std::string("atomic_") + NewStem;
+  Info.PostProc = [=](std::vector<Value *> &Ops){
+    for (auto &I:PostOps){
+      Ops.push_back(addInt32(I));
+    }
+  };
+
+  transAtomicBuiltin(CI, Info);
+}
+
+void
+OCL20ToSPIRV::transAtomicBuiltin(CallInst* CI,
+    OCLBuiltinTransInfo& Info) {
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  mutateCallInstSPIRV(M, CI, [=](CallInst * CI, std::vector<Value *> &Args){
+    Info.PostProc(Args);
+    // Order of args in OCL20:
+    // object, 0-2 other args, 1-2 order, scope
+    const size_t NumOrder = getAtomicBuiltinNumMemoryOrderArgs(Info.UniqName);
+    const size_t ArgsCount = Args.size();
+    const size_t ScopeIdx = ArgsCount - 1;
+    const size_t OrderIdx = ScopeIdx - NumOrder;
+    Args[ScopeIdx] = mapUInt(M, cast<ConstantInt>(Args[ScopeIdx]),
+        [](unsigned I){
+      return map<Scope>(static_cast<OCLScopeKind>(I));
+    });
+    for (size_t I = 0; I < NumOrder; ++I)
+      Args[OrderIdx + I] = mapUInt(M, cast<ConstantInt>(Args[OrderIdx + I]),
+          [](unsigned Ord) {
+      return mapOCLMemSemanticToSPIRV(0, static_cast<OCLMemOrderKind>(Ord));
+    });
+    // Order of args in SPIR-V:
+    // object, scope, 1-2 order, 0-2 other args
+    std::swap(Args[1], Args[ScopeIdx]);
+    if(OrderIdx > 2) {
+      // For atomic_compare_exchange the swap above puts Comparator/Expected
+      // argument just where it should be, so don't move the last argument then.
+      int offset = Info.UniqName.find("atomic_compare_exchange") == 0 ? 1 : 0;
+      std::rotate(Args.begin() + 2, Args.begin() + OrderIdx,
+                  Args.end() - offset);
+    }
+    return getSPIRVFuncName(OCLSPIRVBuiltinMap::map(Info.UniqName));
+  }, &Attrs);
+}
+
+void
+OCL20ToSPIRV::visitCallWorkGroupBarrier(CallInst* CI) {
+  auto Lit = getWorkGroupBarrierLiterals(CI);
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  mutateCallInstSPIRV(M, CI, [=](CallInst *, std::vector<Value *> &Args){
+    Args.resize(3);
+    Args[0] = addInt32(map<Scope>(std::get<2>(Lit)));
+    Args[1] = addInt32(map<Scope>(std::get<1>(Lit)));
+    Args[2] = addInt32(mapOCLMemFenceFlagToSPIRV(std::get<0>(Lit)));
+    return getSPIRVFuncName(OpControlBarrier);
+  }, &Attrs);
+}
+
+void OCL20ToSPIRV::visitCallConvert(CallInst* CI,
+    StringRef MangledName, const std::string& DemangledName) {
+  if (eraseUselessConvert(CI, MangledName, DemangledName))
+    return;
+  Op OC = OpNop;
+  auto TargetTy = CI->getType();
+  auto SrcTy = CI->getArgOperand(0)->getType();
+  if (isa<VectorType>(TargetTy))
+    TargetTy = TargetTy->getVectorElementType();
+  if (isa<VectorType>(SrcTy))
+    SrcTy = SrcTy->getVectorElementType();
+  auto IsTargetInt = isa<IntegerType>(TargetTy);
+
+  std::string TargetTyName = DemangledName.substr(
+      strlen(kOCLBuiltinName::ConvertPrefix));
+  auto FirstUnderscoreLoc = TargetTyName.find('_');
+  if (FirstUnderscoreLoc != std::string::npos)
+    TargetTyName = TargetTyName.substr(0, FirstUnderscoreLoc);
+  TargetTyName = std::string("_R") + TargetTyName;
+
+  std::string Sat = DemangledName.find("_sat") != std::string::npos ?
+      "_sat" : "";
+  auto TargetSigned = DemangledName[8] != 'u';
+  if (isa<IntegerType>(SrcTy)) {
+    bool Signed = isLastFuncParamSigned(MangledName);
+    if (IsTargetInt) {
+      if (!Sat.empty() && TargetSigned != Signed) {
+        OC = Signed ? OpSatConvertSToU : OpSatConvertUToS;
+        Sat = "";
+      } else
+        OC = Signed ? OpSConvert : OpUConvert;
+    } else
+      OC = Signed ? OpConvertSToF : OpConvertUToF;
+  } else {
+    if (IsTargetInt) {
+      OC = TargetSigned ? OpConvertFToS : OpConvertFToU;
+    } else
+      OC = OpFConvert;
+  }
+  auto Loc = DemangledName.find("_rt");
+  std::string Rounding;
+  if (Loc != std::string::npos &&
+      !(isa<IntegerType>(SrcTy) && IsTargetInt)) {
+    Rounding = DemangledName.substr(Loc, 4);
+  }
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  mutateCallInstSPIRV(M, CI, [=](CallInst *, std::vector<Value *> &Args){
+    return getSPIRVFuncName(OC, TargetTyName + Sat + Rounding);
+  }, &Attrs);
+}
+
+void OCL20ToSPIRV::visitCallGroupBuiltin(CallInst* CI,
+    StringRef MangledName, const std::string& OrigDemangledName) {
+  auto F = CI->getCalledFunction();
+  std::vector<int> PreOps;
+  std::string DemangledName = OrigDemangledName;
+
+  if (DemangledName == kOCLBuiltinName::WorkGroupBarrier)
+    return;
+  if (DemangledName == kOCLBuiltinName::WaitGroupEvent) {
+    PreOps.push_back(ScopeWorkgroup);
+  } else if (DemangledName.find(kOCLBuiltinName::WorkGroupPrefix) == 0) {
+    DemangledName.erase(0, strlen(kOCLBuiltinName::WorkPrefix));
+    PreOps.push_back(ScopeWorkgroup);
+  } else if (DemangledName.find(kOCLBuiltinName::SubGroupPrefix) == 0) {
+    DemangledName.erase(0, strlen(kOCLBuiltinName::SubPrefix));
+    PreOps.push_back(ScopeSubgroup);
+  } else
+    return;
+
+  if (DemangledName != kOCLBuiltinName::WaitGroupEvent) {
+    StringRef GroupOp = DemangledName;
+    GroupOp = GroupOp.drop_front(strlen(kSPIRVName::GroupPrefix));
+    SPIRSPIRVGroupOperationMap::foreach_conditional([&](const std::string &S,
+        SPIRVGroupOperationKind G){
+      if (!GroupOp.startswith(S))
+        return true; // continue
+      PreOps.push_back(G);
+      StringRef Op = GroupOp.drop_front(S.size() + 1);
+      assert(!Op.empty() && "Invalid OpenCL group builtin function");
+      char OpTyC = 0;
+      auto NeedSign = Op == "max" || Op == "min";
+      auto OpTy = F->getReturnType();
+      if (OpTy->isFloatingPointTy())
+        OpTyC = 'f';
+      else if (OpTy->isIntegerTy()) {
+        if (!NeedSign)
+          OpTyC = 'i';
+        else {
+          if (isLastFuncParamSigned(F->getName()))
+            OpTyC = 's';
+          else
+            OpTyC = 'u';
+        }
+      } else
+        llvm_unreachable("Invalid OpenCL group builtin argument type");
+
+      DemangledName = std::string(kSPIRVName::GroupPrefix) + OpTyC + Op.str();
+      return false; // break out of loop
+    });
+  }
+
+  bool IsGroupAllAny = (DemangledName.find("_all") != std::string::npos ||
+                        DemangledName.find("_any") != std::string::npos);
+
+  auto Consts = getInt32(M, PreOps);
+  OCLBuiltinTransInfo Info;
+  if (IsGroupAllAny)
+    Info.RetTy = Type::getInt1Ty(*Ctx);
+  Info.UniqName = DemangledName;
+  Info.PostProc = [=](std::vector<Value *> &Ops) {
+    if (IsGroupAllAny) {
+      IRBuilder<> IRB(CI);
+      Ops[0] =
+          IRB.CreateICmpNE(Ops[0], ConstantInt::get(Type::getInt32Ty(*Ctx), 0));
+    }
+    size_t E = Ops.size();
+    if (DemangledName == "group_broadcast" && E > 2) {
+      assert(E == 3 || E == 4);
+      makeVector(CI, Ops, std::make_pair(Ops.begin() + 1, Ops.end()));
+    }
+    Ops.insert(Ops.begin(), Consts.begin(), Consts.end());
+  };
+  transBuiltin(CI, Info);
+}
+
+void
+OCL20ToSPIRV::transBuiltin(CallInst* CI,
+    OCLBuiltinTransInfo& Info) {
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  Op OC = OpNop;
+  unsigned ExtOp = ~0U;
+  if (StringRef(Info.UniqName).startswith(kSPIRVName::Prefix))
+      return;
+  if (OCLSPIRVBuiltinMap::find(Info.UniqName, &OC))
+    Info.UniqName = getSPIRVFuncName(OC);
+  else if ((ExtOp = getExtOp(Info.MangledName, Info.UniqName)) != ~0U)
+    Info.UniqName = getSPIRVExtFuncName(SPIRVEIS_OpenCL, ExtOp);
+  else
+    return;
+  if (!Info.RetTy)
+    mutateCallInstSPIRV(M, CI,
+                        [=](CallInst *, std::vector<Value *> &Args) {
+                          Info.PostProc(Args);
+                          return Info.UniqName + Info.Postfix;
+                        },
+                        &Attrs);
+  else
+    mutateCallInstSPIRV(
+        M, CI,
+        [=](CallInst *, std::vector<Value *> &Args, Type *&RetTy) {
+          Info.PostProc(Args);
+          RetTy = Info.RetTy;
+          return Info.UniqName + Info.Postfix;
+        },
+        [=](CallInst *NewCI) -> Instruction * {
+          if (NewCI->getType()->isIntegerTy() && CI->getType()->isIntegerTy())
+            return CastInst::CreateIntegerCast(NewCI, CI->getType(),
+                                               Info.isRetSigned, "", CI);
+          else
+            return CastInst::CreatePointerBitCastOrAddrSpaceCast(
+                NewCI, CI->getType(), "", CI);
+        },
+        &Attrs);
+}
+
+void
+OCL20ToSPIRV::visitCallPipeBuiltin(CallInst* CI,
+    StringRef MangledName, const std::string& DemangledName) {
+  std::string NewName = DemangledName;
+  // Transform OpenCL read_pipe/write_pipe builtin function names
+  // with reserve_id argument to reserved_read_pipe/reserved_write_pipe.
+  if ((DemangledName.find(kOCLBuiltinName::ReadPipe) == 0 ||
+      DemangledName.find(kOCLBuiltinName::WritePipe) == 0)
+      && CI->getNumArgOperands() > 4)
+    NewName = std::string(kSPIRVName::ReservedPrefix) + DemangledName;
+  OCLBuiltinTransInfo Info;
+  Info.UniqName = NewName;
+  transBuiltin(CI, Info);
+}
+
+void OCL20ToSPIRV::visitCallReadImageMSAA(CallInst *CI, StringRef MangledName,
+                                          const std::string &DemangledName) {
+  assert(MangledName.find("msaa") != StringRef::npos);
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  mutateCallInstSPIRV(
+      M, CI,
+      [=](CallInst *, std::vector<Value *> &Args) {
+        Args.insert(Args.begin() + 2, getInt32(M, ImageOperandsSampleMask));
+        return getSPIRVFuncName(OpImageRead,
+                                std::string(kSPIRVPostfix::ExtDivider) +
+                                    getPostfixForReturnType(CI));
+      },
+      &Attrs);
+}
+
+void OCL20ToSPIRV::visitCallReadImageWithSampler(
+    CallInst *CI, StringRef MangledName, const std::string &DemangledName) {
+  assert (MangledName.find(kMangledName::Sampler) != StringRef::npos);
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  bool isRetScalar = !CI->getType()->isVectorTy();
+  mutateCallInstSPIRV(
+      M, CI,
+      [=](CallInst *, std::vector<Value *> &Args, Type *&Ret) {
+        auto ImageTy = getAnalysis<OCLTypeToSPIRV>().getAdaptedType(Args[0]);
+        if (isOCLImageType(ImageTy))
+          ImageTy = getSPIRVImageTypeFromOCL(M, ImageTy);
+        auto SampledImgTy = getSPIRVTypeByChangeBaseTypeName(
+            M, ImageTy, kSPIRVTypeName::Image, kSPIRVTypeName::SampledImg);
+        Value *SampledImgArgs[] = {Args[0], Args[1]};
+        auto SampledImg = addCallInstSPIRV(
+            M, getSPIRVFuncName(OpSampledImage), SampledImgTy, SampledImgArgs,
+            nullptr, CI, kSPIRVName::TempSampledImage);
+
+        Args[0] = SampledImg;
+        Args.erase(Args.begin() + 1, Args.begin() + 2);
+
+        switch (Args.size()) {
+        case 2: // no lod
+          Args.push_back(getInt32(M, ImageOperandsMask::ImageOperandsLodMask));
+          Args.push_back(getFloat32(M, 0.f));
+          break;
+        case 3: // explicit lod
+          Args.insert(Args.begin() + 2,
+                      getInt32(M, ImageOperandsMask::ImageOperandsLodMask));
+          break;
+        case 4: // gradient
+          Args.insert(Args.begin() + 2,
+                      getInt32(M, ImageOperandsMask::ImageOperandsGradMask));
+          break;
+        default:
+          assert(0 && "read_image* with unhandled number of args!");
+        }
+
+        // SPIR-V intruction always returns 4-element vector
+        if (isRetScalar)
+          Ret = VectorType::get(Ret, 4);
+        return getSPIRVFuncName(OpImageSampleExplicitLod,
+                                std::string(kSPIRVPostfix::ExtDivider) +
+                                    getPostfixForReturnType(Ret));
+      },
+      [&](CallInst *CI) -> Instruction * {
+        if (isRetScalar)
+          return ExtractElementInst::Create(CI, getSizet(M, 0), "",
+                                            CI->getNextNode());
+        return CI;
+      },
+      &Attrs);
+}
+
+void
+OCL20ToSPIRV::visitCallGetImageSize(CallInst* CI,
+    StringRef MangledName, const std::string& DemangledName) {
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  StringRef TyName;
+  auto IsImg = isOCLImageType(CI->getArgOperand(0)->getType(), &TyName);
+  assert(IsImg);
+  SPIRVTypeImageDescriptor Desc = map<SPIRVTypeImageDescriptor>(TyName.str());
+  unsigned Dim = getImageDimension(Desc.Dim) + Desc.Arrayed;
+  assert(Dim > 0 && "Ivalid image dimention.");
+  mutateCallInstSPIRV(M, CI,
+    [&](CallInst *, std::vector<Value *> &Args, Type *&Ret){
+      assert(Args.size() == 1);
+      Ret = CI->getType()->isIntegerTy(64) ? Type::getInt64Ty(*Ctx)
+                                           : Type::getInt32Ty(*Ctx);
+      if (Dim > 1)
+        Ret = VectorType::get(Ret, Dim);
+      if (Desc.Dim == DimBuffer)
+        return getSPIRVFuncName(OpImageQuerySize, CI->getType());
+      else {
+        Args.push_back(getInt32(M, 0));
+        return getSPIRVFuncName(OpImageQuerySizeLod, CI->getType());
+      }
+    },
+    [&](CallInst *NCI)->Instruction * {
+      if (Dim == 1)
+        return NCI;
+      if (DemangledName == kOCLBuiltinName::GetImageDim) {
+        if (Desc.Dim == Dim3D) {
+          auto ZeroVec = ConstantVector::getSplat(3,
+            Constant::getNullValue(NCI->getType()->getVectorElementType()));
+          Constant *Index[] = {getInt32(M, 0), getInt32(M, 1),
+              getInt32(M, 2), getInt32(M, 3)};
+          return new ShuffleVectorInst(NCI, ZeroVec,
+             ConstantVector::get(Index), "", CI);
+
+        } else if (Desc.Dim == Dim2D && Desc.Arrayed) {
+          Constant *Index[] = {getInt32(M, 0), getInt32(M, 1)};
+          Constant *mask = ConstantVector::get(Index);
+          return new ShuffleVectorInst(NCI, UndefValue::get(NCI->getType()),
+                                       mask, NCI->getName(), CI);
+        }
+        return NCI;
+      }
+      unsigned I = StringSwitch<unsigned>(DemangledName)
+          .Case(kOCLBuiltinName::GetImageWidth, 0)
+          .Case(kOCLBuiltinName::GetImageHeight, 1)
+          .Case(kOCLBuiltinName::GetImageDepth, 2)
+          .Case(kOCLBuiltinName::GetImageArraySize, Dim - 1);
+      return ExtractElementInst::Create(NCI, getUInt32(M, I), "",
+          NCI->getNextNode());
+    },
+  &Attrs);
+}
+
+/// Remove trivial conversion functions
+bool
+OCL20ToSPIRV::eraseUselessConvert(CallInst *CI,
+    const std::string &MangledName,
+    const std::string &DemangledName) {
+  auto TargetTy = CI->getType();
+  auto SrcTy = CI->getArgOperand(0)->getType();
+  if (isa<VectorType>(TargetTy))
+    TargetTy = TargetTy->getVectorElementType();
+  if (isa<VectorType>(SrcTy))
+    SrcTy = SrcTy->getVectorElementType();
+  if (TargetTy == SrcTy) {
+    if (isa<IntegerType>(TargetTy) &&
+        DemangledName.find("_sat") != std::string::npos &&
+        isLastFuncParamSigned(MangledName) != (DemangledName[8] != 'u'))
+      return false;
+    CI->getArgOperand(0)->takeName(CI);
+    SPIRVDBG(dbgs() << "[regularizeOCLConvert] " << *CI << " <- " <<
+        *CI->getArgOperand(0) << '\n');
+    CI->replaceAllUsesWith(CI->getArgOperand(0));
+    ValuesToDelete.insert(CI);
+    ValuesToDelete.insert(CI->getCalledFunction());
+    return true;
+  }
+  return false;
+}
+
+void
+OCL20ToSPIRV::visitCallBuiltinSimple(CallInst* CI,
+    StringRef MangledName, const std::string& DemangledName) {
+  OCLBuiltinTransInfo Info;
+  Info.MangledName = MangledName.str();
+  Info.UniqName = DemangledName;
+  transBuiltin(CI, Info);
+}
+
+/// Translates OCL work-item builtin functions to SPIRV builtin variables.
+/// Function like get_global_id(i) -> x = load GlobalInvocationId; extract x, i
+/// Function like get_work_dim() -> load WorkDim
+void OCL20ToSPIRV::transWorkItemBuiltinsToVariables() {
+  DEBUG(dbgs() << "Enter transWorkItemBuiltinsToVariables\n");
+  std::vector<Function *> WorkList;
+  for (auto I = M->begin(), E = M->end(); I != E; ++I) {
+    std::string DemangledName;
+    if (!oclIsBuiltin(I->getName(), &DemangledName))
+      continue;
+    DEBUG(dbgs() << "Function demangled name: " << DemangledName << '\n');
+    std::string BuiltinVarName;
+    SPIRVBuiltinVariableKind BVKind;
+    if (!SPIRSPIRVBuiltinVariableMap::find(DemangledName, &BVKind))
+      continue;
+    BuiltinVarName = std::string(kSPIRVName::Prefix) +
+        SPIRVBuiltInNameMap::map(BVKind);
+    DEBUG(dbgs() << "builtin variable name: " << BuiltinVarName << '\n');
+    bool IsVec = I->getFunctionType()->getNumParams() > 0;
+    Type *GVType = IsVec ? VectorType::get(I->getReturnType(),3) :
+        I->getReturnType();
+    auto BV = new GlobalVariable(*M, GVType,
+        true,
+        GlobalValue::ExternalLinkage,
+        nullptr, BuiltinVarName,
+        0,
+        GlobalVariable::NotThreadLocal,
+        SPIRAS_Constant);
+    std::vector<Instruction *> InstList;
+    for (auto UI = I->user_begin(), UE = I->user_end(); UI != UE; ++UI) {
+      auto CI = dyn_cast<CallInst>(*UI);
+      assert(CI && "invalid instruction");
+      Value * NewValue = new LoadInst(BV, "", CI);
+      DEBUG(dbgs() << "Transform: " << *CI << " => " << *NewValue << '\n');
+      if (IsVec) {
+        NewValue = ExtractElementInst::Create(NewValue,
+          CI->getArgOperand(0),
+          "", CI);
+        DEBUG(dbgs() << *NewValue << '\n');
+      }
+      NewValue->takeName(CI);
+      CI->replaceAllUsesWith(NewValue);
+      InstList.push_back(CI);
+    }
+    for (auto &Inst:InstList) {
+      Inst->dropAllReferences();
+      Inst->removeFromParent();
+    }
+    WorkList.push_back(&(*I));
+  }
+  for (auto &I:WorkList) {
+    I->dropAllReferences();
+    I->removeFromParent();
+  }
+}
+
+void
+OCL20ToSPIRV::visitCallReadWriteImage(CallInst* CI,
+    StringRef MangledName, const std::string& DemangledName) {
+  OCLBuiltinTransInfo Info;
+  if (DemangledName.find(kOCLBuiltinName::ReadImage) == 0)
+    Info.UniqName = kOCLBuiltinName::ReadImage;
+
+  if (DemangledName.find(kOCLBuiltinName::WriteImage) == 0)
+  {
+    Info.UniqName = kOCLBuiltinName::WriteImage;
+    Info.PostProc = [&](std::vector<Value*> &Args) {
+        if (Args.size() == 4) // write with lod
+        {
+            auto Lod = Args[2];
+            Args.erase(Args.begin() + 2);
+            Args.push_back(getInt32(M, ImageOperandsMask::ImageOperandsLodMask));
+            Args.push_back(Lod);
+        }
+    };
+  }
+
+  transBuiltin(CI, Info);
+}
+
+void
+OCL20ToSPIRV::visitCallToAddr(CallInst* CI, StringRef MangledName,
+    const std::string &DemangledName) {
+  auto AddrSpace = static_cast<SPIRAddressSpace>(
+      CI->getType()->getPointerAddressSpace());
+  OCLBuiltinTransInfo Info;
+  Info.UniqName = DemangledName;
+  Info.Postfix = std::string(kSPIRVPostfix::Divider) + "To" +
+      SPIRAddrSpaceCapitalizedNameMap::map(AddrSpace);
+  auto StorageClass = addInt32(SPIRSPIRVAddrSpaceMap::map(AddrSpace));
+  Info.RetTy = getInt8PtrTy(cast<PointerType>(CI->getType()));
+  Info.PostProc = [=](std::vector<Value *> &Ops){
+    auto P = Ops.back();
+    Ops.pop_back();
+    Ops.push_back(castToInt8Ptr(P, CI));
+    Ops.push_back(StorageClass);
+  };
+  transBuiltin(CI, Info);
+}
+
+void OCL20ToSPIRV::visitCallRelational(CallInst *CI,
+                                       const std::string &DemangledName) {
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  Op OC = OpNop;
+  OCLSPIRVBuiltinMap::find(DemangledName, &OC);
+  std::string SPIRVName = getSPIRVFuncName(OC);
+  mutateCallInstSPIRV(
+      M, CI,
+      [=](CallInst *, std::vector<Value *> &Args, Type *&Ret) {
+        Ret = Type::getInt1Ty(*Ctx);
+        if (CI->getOperand(0)->getType()->isVectorTy())
+          Ret = VectorType::get(
+              Type::getInt1Ty(*Ctx),
+              CI->getOperand(0)->getType()->getVectorNumElements());
+        return SPIRVName;
+      },
+      [=](CallInst *NewCI) -> Instruction * {
+        Value *False = nullptr, *True = nullptr;
+        if (NewCI->getType()->isVectorTy()) {
+          Type *IntTy = Type::getInt32Ty(*Ctx);
+          if (cast<VectorType>(NewCI->getOperand(0)->getType())
+                  ->getElementType()
+                  ->isDoubleTy())
+            IntTy = Type::getInt64Ty(*Ctx);
+          if (cast<VectorType>(NewCI->getOperand(0)->getType())
+                  ->getElementType()
+                  ->isHalfTy())
+            IntTy = Type::getInt16Ty(*Ctx);
+          Type *VTy = VectorType::get(IntTy,
+                                      NewCI->getType()->getVectorNumElements());
+          False = Constant::getNullValue(VTy);
+          True = Constant::getAllOnesValue(VTy);
+        } else {
+          False = getInt32(M, 0);
+          True = getInt32(M, 1);
+        }
+        return SelectInst::Create(NewCI, True, False, "", NewCI->getNextNode());
+      },
+      &Attrs);
+}
+
+void
+OCL20ToSPIRV::visitCallVecLoadStore(CallInst* CI,
+    StringRef MangledName, const std::string& OrigDemangledName) {
+  std::vector<int> PreOps;
+  std::string DemangledName = OrigDemangledName;
+  if (DemangledName.find(kOCLBuiltinName::VLoadPrefix) == 0 &&
+      DemangledName != kOCLBuiltinName::VLoadHalf) {
+    SPIRVWord Width = getVecLoadWidth(DemangledName);
+    SPIRVDBG(spvdbgs() << "[visitCallVecLoadStore] DemangledName: " <<
+        DemangledName << " Width: " << Width << '\n');
+    PreOps.push_back(Width);
+  } else if (DemangledName.find(kOCLBuiltinName::RoundingPrefix)
+      != std::string::npos) {
+    auto R = SPIRSPIRVFPRoundingModeMap::map(DemangledName.substr(
+        DemangledName.find(kOCLBuiltinName::RoundingPrefix) + 1, 3));
+    PreOps.push_back(R);
+  }
+
+  if (DemangledName.find(kOCLBuiltinName::VLoadAPrefix) == 0)
+    transVecLoadStoreName(DemangledName, kOCLBuiltinName::VLoadAPrefix, true);
+  else
+    transVecLoadStoreName(DemangledName, kOCLBuiltinName::VLoadPrefix, false);
+
+  if (DemangledName.find(kOCLBuiltinName::VStoreAPrefix) == 0)
+    transVecLoadStoreName(DemangledName, kOCLBuiltinName::VStoreAPrefix, true);
+  else
+    transVecLoadStoreName(DemangledName, kOCLBuiltinName::VStorePrefix, false);
+
+
+  auto Consts = getInt32(M, PreOps);
+  OCLBuiltinTransInfo Info;
+  Info.MangledName = MangledName;
+  Info.UniqName = DemangledName;
+  if (DemangledName.find(kOCLBuiltinName::VLoadPrefix) == 0)
+    Info.Postfix = std::string(kSPIRVPostfix::ExtDivider) +
+      getPostfixForReturnType(CI);
+  Info.PostProc = [=](std::vector<Value *> &Ops){
+    Ops.insert(Ops.end(), Consts.begin(), Consts.end());
+  };
+  transBuiltin(CI, Info);
+}
+
+void OCL20ToSPIRV::visitCallGetFence(CallInst *CI, StringRef MangledName,
+                                     const std::string &DemangledName) {
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  Op OC = OpNop;
+  OCLSPIRVBuiltinMap::find(DemangledName, &OC);
+  std::string SPIRVName = getSPIRVFuncName(OC);
+  mutateCallInstSPIRV(M, CI, [=](CallInst *, std::vector<Value *> &Args,
+                                 Type *&Ret) { return SPIRVName; },
+            [=](CallInst *NewCI) -> Instruction * {
+              return BinaryOperator::CreateLShr(NewCI, getInt32(M, 8), "", CI);
+            },
+            &Attrs);
+}
+
+void OCL20ToSPIRV::visitCallDot(CallInst *CI) {
+  IRBuilder<> Builder(CI);
+  Value *FMulVal = Builder.CreateFMul(CI->getOperand(0), CI->getOperand(1));
+  CI->replaceAllUsesWith(FMulVal);
+  CI->dropAllReferences();
+  CI->removeFromParent();
+}
+
+void OCL20ToSPIRV::visitCallScalToVec(CallInst *CI, StringRef MangledName,
+                                      const std::string &DemangledName) {
+  // Check if all arguments have the same type - it's simple case.
+  auto Uniform = true;
+  auto IsArg0Vector = isa<VectorType>(CI->getOperand(0)->getType());
+  for (unsigned I = 1, E = CI->getNumArgOperands(); Uniform && (I != E); ++I) {
+    Uniform = isa<VectorType>(CI->getOperand(I)->getType()) == IsArg0Vector;
+  }
+  if (Uniform) {
+    visitCallBuiltinSimple(CI, MangledName, DemangledName);
+    return;
+  }
+
+  std::vector<unsigned int> VecPos;
+  std::vector<unsigned int> ScalarPos;
+  if (DemangledName == kOCLBuiltinName::FMin ||
+      DemangledName == kOCLBuiltinName::FMax ||
+      DemangledName == kOCLBuiltinName::Min ||
+      DemangledName == kOCLBuiltinName::Max) {
+    VecPos.push_back(0);
+    ScalarPos.push_back(1);
+  } else if (DemangledName == kOCLBuiltinName::Clamp) {
+    VecPos.push_back(0);
+    ScalarPos.push_back(1);
+    ScalarPos.push_back(2);
+  } else if (DemangledName == kOCLBuiltinName::Mix) {
+    VecPos.push_back(0);
+    VecPos.push_back(1);
+    ScalarPos.push_back(2);
+  } else if (DemangledName == kOCLBuiltinName::Step) {
+    VecPos.push_back(1);
+    ScalarPos.push_back(0);
+  } else if (DemangledName == kOCLBuiltinName::SmoothStep) {
+    VecPos.push_back(2);
+    ScalarPos.push_back(0);
+    ScalarPos.push_back(1);
+  }
+
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  mutateCallInstSPIRV(
+      M, CI,
+      [=](CallInst *, std::vector<Value *> &Args) {
+        Args.resize(VecPos.size() + ScalarPos.size());
+        for (auto I : VecPos) {
+          Args[I] = CI->getOperand(I);
+        }
+        auto VecArgWidth =
+            CI->getOperand(VecPos[0])->getType()->getVectorNumElements();
+        for (auto I : ScalarPos) {
+          Instruction *Inst = InsertElementInst::Create(
+              UndefValue::get(CI->getOperand(VecPos[0])->getType()),
+              CI->getOperand(I), getInt32(M, 0), "", CI);
+          Value *NewVec = new ShuffleVectorInst(
+              Inst, UndefValue::get(CI->getOperand(VecPos[0])->getType()),
+              ConstantVector::getSplat(VecArgWidth, getInt32(M, 0)), "", CI);
+
+          Args[I] = NewVec;
+        }
+        return getSPIRVExtFuncName(SPIRVEIS_OpenCL,
+                                   getExtOp(MangledName, DemangledName));
+      },
+      &Attrs);
+}
+
+void OCL20ToSPIRV::visitCallGetImageChannel(CallInst *CI, StringRef MangledName,
+                                            const std::string &DemangledName,
+                                            unsigned int Offset) {
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  Op OC = OpNop;
+  OCLSPIRVBuiltinMap::find(DemangledName, &OC);
+  std::string SPIRVName = getSPIRVFuncName(OC);
+  mutateCallInstSPIRV(M, CI, [=](CallInst *, std::vector<Value *> &Args,
+                                 Type *&Ret) { return SPIRVName; },
+                      [=](CallInst *NewCI) -> Instruction * {
+                        return BinaryOperator::CreateAdd(
+                            NewCI, getInt32(M, Offset), "", CI);
+                      },
+                      &Attrs);
+}
+}
+
+INITIALIZE_PASS_BEGIN(OCL20ToSPIRV, "cl20tospv", "Transform OCL 2.0 to SPIR-V",
+    false, false)
+INITIALIZE_PASS_DEPENDENCY(OCLTypeToSPIRV)
+INITIALIZE_PASS_END(OCL20ToSPIRV, "cl20tospv", "Transform OCL 2.0 to SPIR-V",
+    false, false)
+
+ModulePass *llvm::createOCL20ToSPIRV() {
+  return new OCL20ToSPIRV();
+}
diff --git a/OCLTypeToSPIRV.cpp b/OCLTypeToSPIRV.cpp
index be58ff1..5784197 100644
--- a/OCLTypeToSPIRV.cpp
+++ b/OCLTypeToSPIRV.cpp
@@ -125,7 +125,7 @@ static unsigned
 getArgIndex(Function *F, Value *V) {
   auto A = F->arg_begin(), E = F->arg_end();
   for (unsigned I = 0; A != E; ++I, ++A) {
-    if (A == V)
+    if (&*A == V)
       return I;
   }
   llvm_unreachable("Not argument of function");
@@ -136,7 +136,7 @@ static Argument*
 getArg(Function *F, unsigned I) {
   auto AI = F->arg_begin();
   std::advance(AI, I);
-  return AI;
+  return &*AI;
 }
 
 /// Create a new function type if \param F has arguments in AdaptedTy, and
@@ -291,7 +291,7 @@ OCLTypeToSPIRV::adaptArgumentsByMetadata(Function* F) {
     auto OCLTyStr = getMDOperandAsString(TypeMD, I);
     auto NewTy = *PI;
     if (OCLTyStr == OCL_TYPE_NAME_SAMPLER_T && !NewTy->isStructTy()) {
-      addAdaptedType(Arg, getSamplerType(M));
+      addAdaptedType(&*Arg, getSamplerType(M));
       Changed = true;
     } else if (isPointerToOpaqueStructType(NewTy)) {
       auto STName = NewTy->getPointerElementType()->getStructName();
@@ -301,7 +301,7 @@ OCLTypeToSPIRV::adaptArgumentsByMetadata(Function* F) {
         auto AccMD = getArgAccessQualifierMetadata(F);
         assert(AccMD && "Invalid access qualifier metadata");
         auto AccStr = getMDOperandAsString(AccMD, I);
-        addAdaptedType(Arg, getOrCreateOpaquePtrType(M,
+        addAdaptedType(&*Arg, getOrCreateOpaquePtrType(M,
             mapOCLTypeNameToSPIRV(Ty, AccStr)));
         Changed = true;
       }
diff --git a/SPIRVLowerConstExpr.cpp b/SPIRVLowerConstExpr.cpp
index 82796a0..1d9b58f 100644
--- a/SPIRVLowerConstExpr.cpp
+++ b/SPIRVLowerConstExpr.cpp
@@ -120,7 +120,7 @@ SPIRVLowerConstExpr::visit(Module *M) {
       auto FBegin = I->begin();
       for (auto BI = FBegin, BE = I->end(); BI != BE; ++BI) {
         for (auto II = BI->begin(), IE = BI->end(); II != IE; ++II) {
-          WorkList.push_back(II);
+          WorkList.push_back(&*II);
         }
       }
       while (!WorkList.empty()) {
@@ -132,7 +132,7 @@ SPIRVLowerConstExpr::visit(Module *M) {
           if (auto CE = dyn_cast<ConstantExpr>(Op)) {
             SPIRVDBG(dbgs() << "[lowerConstantExpressions] " << *CE;)
             auto ReplInst = CE->getAsInstruction();
-            ReplInst->insertBefore(FBegin->begin());
+            ReplInst->insertBefore(&*FBegin->begin());
             SPIRVDBG(dbgs() << " -> " << *ReplInst << '\n';)
             WorkList.push_front(ReplInst);
             std::vector<Instruction *> Users;
@@ -141,7 +141,7 @@ SPIRVLowerConstExpr::visit(Module *M) {
               SPIRVDBG(dbgs() << "[lowerConstantExpressions] Use: " <<
                   *U << '\n';)
               if (auto InstUser = dyn_cast<Instruction>(U)) {
-                if (InstUser->getParent()->getParent() != I)
+                if (InstUser->getParent()->getParent() != &*I)
                   continue;
                 Users.push_back(InstUser);
               }
diff --git a/SPIRVLowerOCLBlocks.cpp b/SPIRVLowerOCLBlocks.cpp
index 24405cd..2cf9aff 100644
--- a/SPIRVLowerOCLBlocks.cpp
+++ b/SPIRVLowerOCLBlocks.cpp
@@ -1,471 +1,476 @@
-//===- SPIRVLowerOCLBlocks.cpp - Lower OpenCL blocks ------------*- C++ -*-===//
-//
-//                     The LLVM/SPIR-V Translator
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-// Copyright (c) 2014 Advanced Micro Devices, Inc. All rights reserved.
-//
-// Permission is hereby granted, free of charge, to any person obtaining a
-// copy of this software and associated documentation files (the "Software"),
-// to deal with the Software without restriction, including without limitation
-// the rights to use, copy, modify, merge, publish, distribute, sublicense,
-// and/or sell copies of the Software, and to permit persons to whom the
-// Software is furnished to do so, subject to the following conditions:
-//
-// Redistributions of source code must retain the above copyright notice,
-// this list of conditions and the following disclaimers.
-// Redistributions in binary form must reproduce the above copyright notice,
-// this list of conditions and the following disclaimers in the documentation
-// and/or other materials provided with the distribution.
-// Neither the names of Advanced Micro Devices, Inc., nor the names of its
-// contributors may be used to endorse or promote products derived from this
-// Software without specific prior written permission.
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-// CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
-// THE SOFTWARE.
-//
-//===----------------------------------------------------------------------===//
-/// \file
-///
-/// This file implements lowering of OpenCL blocks to functions.
-///
-//===----------------------------------------------------------------------===//
-
-#ifndef OCLLOWERBLOCKS_H_
-#define OCLLOWERBLOCKS_H_
-
-#include "SPIRVInternal.h"
-#include "OCLUtil.h"
-
-#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/SetVector.h"
-#include "llvm/ADT/StringSwitch.h"
-#include "llvm/ADT/Triple.h"
-#include "llvm/Analysis/AliasAnalysis.h"
-#include "llvm/Analysis/AssumptionCache.h"
-#include "llvm/Analysis/CallGraph.h"
-#include "llvm/IR/Verifier.h"
-#include "llvm/Bitcode/ReaderWriter.h"
-#include "llvm/IR/Constants.h"
-#include "llvm/IR/DerivedTypes.h"
-#include "llvm/IR/Function.h"
-#include "llvm/IR/InstrTypes.h"
-#include "llvm/IR/Instructions.h"
-#include "llvm/IR/Module.h"
-#include "llvm/IR/Operator.h"
-#include "llvm/Pass.h"
-#include "llvm/PassSupport.h"
-#include "llvm/Support/Casting.h"
-#include "llvm/Support/Debug.h"
-#include "llvm/Support/raw_ostream.h"
-#include "llvm/Support/ToolOutputFile.h"
-#include "llvm/Transforms/Utils/Cloning.h"
-
-#include <iostream>
-#include <list>
-#include <memory>
-#include <set>
-#include <sstream>
-#include <vector>
-
-#define DEBUG_TYPE "spvblocks"
-
-using namespace llvm;
-using namespace SPIRV;
-using namespace OCLUtil;
-
-namespace SPIRV{
-
-/// Lower SPIR2 blocks to function calls.
-///
-/// SPIR2 representation of blocks:
-///
-/// block = spir_block_bind(bitcast(block_func), context_len, context_align,
-///   context)
-/// block_func_ptr = bitcast(spir_get_block_invoke(block))
-/// context_ptr = spir_get_block_context(block)
-/// ret = block_func_ptr(context_ptr, args)
-///
-/// Propagates block_func to each spir_get_block_invoke through def-use chain of
-/// spir_block_bind, so that
-/// ret = block_func(context, args)
-class SPIRVLowerOCLBlocks: public ModulePass {
-public:
-  SPIRVLowerOCLBlocks():ModulePass(ID), M(nullptr){
-    initializeSPIRVLowerOCLBlocksPass(*PassRegistry::getPassRegistry());
-  }
-
-  virtual void getAnalysisUsage(AnalysisUsage &AU) const {
-    AU.addRequired<CallGraphWrapperPass>();
-    AU.addRequired<AliasAnalysis>();
-    AU.addRequired<AssumptionCacheTracker>();
-  }
-
-  virtual bool runOnModule(Module &Module) {
-    M = &Module;
-    lowerBlockBind();
-    lowerGetBlockInvoke();
-    lowerGetBlockContext();
-    erase(M->getFunction(SPIR_INTRINSIC_GET_BLOCK_INVOKE));
-    erase(M->getFunction(SPIR_INTRINSIC_GET_BLOCK_CONTEXT));
-    erase(M->getFunction(SPIR_INTRINSIC_BLOCK_BIND));
-    DEBUG(dbgs() << "------- After OCLLowerBlocks ------------\n" <<
-                    *M << '\n');
-    return true;
-  }
-
-  static char ID;
-private:
-  const static int MaxIter = 1000;
-  Module *M;
-
-  bool
-  lowerBlockBind() {
-    auto F = M->getFunction(SPIR_INTRINSIC_BLOCK_BIND);
-    if (!F)
-      return false;
-    int Iter = MaxIter;
-    while(lowerBlockBind(F) && Iter > 0){
-      Iter--;
-      DEBUG(dbgs() << "-------------- after iteration " << MaxIter - Iter <<
-          " --------------\n" << *M << '\n');
-    }
-    assert(Iter > 0 && "Too many iterations");
-    return true;
-  }
-
-  bool
-  eraseUselessFunctions() {
-    bool changed = false;
-    for (auto I = M->begin(), E = M->end(); I != E;) {
-      Function *F = I++;
-      if (!GlobalValue::isInternalLinkage(F->getLinkage()) &&
-          !F->isDeclaration())
-        continue;
-
-      dumpUsers(F, "[eraseUselessFunctions] ");
-      for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE;) {
-        auto U = *UI++;
-        if (auto CE = dyn_cast<ConstantExpr>(U)){
-          if (CE->use_empty()) {
-            CE->dropAllReferences();
-            changed = true;
-          }
-        }
-      }
-      if (F->use_empty()) {
-        erase(F);
-        changed = true;
-      }
-    }
-    return changed;
-  }
-
-  void
-  lowerGetBlockInvoke() {
-    if (auto F = M->getFunction(SPIR_INTRINSIC_GET_BLOCK_INVOKE)) {
-      for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE;) {
-        auto CI = dyn_cast<CallInst>(*UI++);
-        assert(CI && "Invalid usage of spir_get_block_invoke");
-        lowerGetBlockInvoke(CI);
-      }
-    }
-  }
-
-  void
-  lowerGetBlockContext() {
-    if (auto F = M->getFunction(SPIR_INTRINSIC_GET_BLOCK_CONTEXT)) {
-      for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE;) {
-        auto CI = dyn_cast<CallInst>(*UI++);
-        assert(CI && "Invalid usage of spir_get_block_context");
-        lowerGetBlockContext(CI);
-      }
-    }
-  }
-  /// Lower calls of spir_block_bind.
-  /// Return true if the Module is changed.
-  bool
-  lowerBlockBind(Function *BlockBindFunc) {
-    bool changed = false;
-    for (auto I = BlockBindFunc->user_begin(), E = BlockBindFunc->user_end();
-        I != E;) {
-      DEBUG(dbgs() << "[lowerBlockBind] " << **I << '\n');
-      // Handle spir_block_bind(bitcast(block_func), context_len,
-      // context_align, context)
-      auto CallBlkBind = cast<CallInst>(*I++);
-      Function *InvF = nullptr;
-      Value *Ctx = nullptr;
-      Value *CtxLen = nullptr;
-      Value *CtxAlign = nullptr;
-      getBlockInvokeFuncAndContext(CallBlkBind, &InvF, &Ctx, &CtxLen,
-          &CtxAlign);
-      for (auto II = CallBlkBind->user_begin(), EE = CallBlkBind->user_end();
-          II != EE;) {
-        auto BlkUser = *II++;
-        SPIRVDBG(dbgs() << "  Block user: " << *BlkUser << '\n');
-        if (auto Ret = dyn_cast<ReturnInst>(BlkUser)) {
-          bool Inlined = false;
-          changed |= lowerReturnBlock(Ret, CallBlkBind, Inlined);
-          if (Inlined)
-            return true;
-        } else if (auto CI = dyn_cast<CallInst>(BlkUser)){
-          auto CallBindF = CI->getCalledFunction();
-          auto Name = CallBindF->getName();
-          std::string DemangledName;
-          if (Name == SPIR_INTRINSIC_GET_BLOCK_INVOKE) {
-            assert(CI->getArgOperand(0) == CallBlkBind);
-            changed |= lowerGetBlockInvoke(CI, cast<Function>(InvF));
-          } else if (Name == SPIR_INTRINSIC_GET_BLOCK_CONTEXT) {
-            assert(CI->getArgOperand(0) == CallBlkBind);
-            // Handle context_ptr = spir_get_block_context(block)
-            lowerGetBlockContext(CI, Ctx);
-            changed = true;
+//===- SPIRVLowerOCLBlocks.cpp - Lower OpenCL blocks ------------*- C++ -*-===//
+//
+//                     The LLVM/SPIR-V Translator
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+// Copyright (c) 2014 Advanced Micro Devices, Inc. All rights reserved.
+//
+// Permission is hereby granted, free of charge, to any person obtaining a
+// copy of this software and associated documentation files (the "Software"),
+// to deal with the Software without restriction, including without limitation
+// the rights to use, copy, modify, merge, publish, distribute, sublicense,
+// and/or sell copies of the Software, and to permit persons to whom the
+// Software is furnished to do so, subject to the following conditions:
+//
+// Redistributions of source code must retain the above copyright notice,
+// this list of conditions and the following disclaimers.
+// Redistributions in binary form must reproduce the above copyright notice,
+// this list of conditions and the following disclaimers in the documentation
+// and/or other materials provided with the distribution.
+// Neither the names of Advanced Micro Devices, Inc., nor the names of its
+// contributors may be used to endorse or promote products derived from this
+// Software without specific prior written permission.
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+// CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+// THE SOFTWARE.
+//
+//===----------------------------------------------------------------------===//
+/// \file
+///
+/// This file implements lowering of OpenCL blocks to functions.
+///
+//===----------------------------------------------------------------------===//
+
+#ifndef OCLLOWERBLOCKS_H_
+#define OCLLOWERBLOCKS_H_
+
+#include "SPIRVInternal.h"
+#include "OCLUtil.h"
+
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SetVector.h"
+#include "llvm/ADT/StringSwitch.h"
+#include "llvm/ADT/Triple.h"
+#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/AssumptionCache.h"
+#include "llvm/Analysis/CallGraph.h"
+#include "llvm/IR/Verifier.h"
+#include "llvm/Bitcode/BitcodeWriter.h"
+#include "llvm/Bitcode/BitcodeReader.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/InstrTypes.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/Operator.h"
+#include "llvm/Pass.h"
+#include "llvm/PassSupport.h"
+#include "llvm/Support/Casting.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Support/ToolOutputFile.h"
+#include "llvm/Transforms/Utils/Cloning.h"
+
+#include <iostream>
+#include <list>
+#include <memory>
+#include <set>
+#include <sstream>
+#include <vector>
+
+#define DEBUG_TYPE "spvblocks"
+
+using namespace llvm;
+using namespace SPIRV;
+using namespace OCLUtil;
+
+namespace SPIRV{
+
+/// Lower SPIR2 blocks to function calls.
+///
+/// SPIR2 representation of blocks:
+///
+/// block = spir_block_bind(bitcast(block_func), context_len, context_align,
+///   context)
+/// block_func_ptr = bitcast(spir_get_block_invoke(block))
+/// context_ptr = spir_get_block_context(block)
+/// ret = block_func_ptr(context_ptr, args)
+///
+/// Propagates block_func to each spir_get_block_invoke through def-use chain of
+/// spir_block_bind, so that
+/// ret = block_func(context, args)
+class SPIRVLowerOCLBlocks: public ModulePass {
+public:
+  SPIRVLowerOCLBlocks():ModulePass(ID), M(nullptr){
+    initializeSPIRVLowerOCLBlocksPass(*PassRegistry::getPassRegistry());
+  }
+
+  virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+    AU.addRequired<CallGraphWrapperPass>();
+    //AU.addRequired<AliasAnalysis>();
+    AU.addRequired<AssumptionCacheTracker>();
+  }
+
+  virtual bool runOnModule(Module &Module) {
+    M = &Module;
+    lowerBlockBind();
+    lowerGetBlockInvoke();
+    lowerGetBlockContext();
+    erase(M->getFunction(SPIR_INTRINSIC_GET_BLOCK_INVOKE));
+    erase(M->getFunction(SPIR_INTRINSIC_GET_BLOCK_CONTEXT));
+    erase(M->getFunction(SPIR_INTRINSIC_BLOCK_BIND));
+    DEBUG(dbgs() << "------- After OCLLowerBlocks ------------\n" <<
+                    *M << '\n');
+    return true;
+  }
+
+  static char ID;
+private:
+  const static int MaxIter = 1000;
+  Module *M;
+
+  bool
+  lowerBlockBind() {
+    auto F = M->getFunction(SPIR_INTRINSIC_BLOCK_BIND);
+    if (!F)
+      return false;
+    int Iter = MaxIter;
+    while(lowerBlockBind(F) && Iter > 0){
+      Iter--;
+      DEBUG(dbgs() << "-------------- after iteration " << MaxIter - Iter <<
+          " --------------\n" << *M << '\n');
+    }
+    assert(Iter > 0 && "Too many iterations");
+    return true;
+  }
+
+  bool
+  eraseUselessFunctions() {
+    bool changed = false;
+    for (auto I = M->begin(), E = M->end(); I != E;) {
+      Function *F = &*I++;
+      if (!GlobalValue::isInternalLinkage(F->getLinkage()) &&
+          !F->isDeclaration())
+        continue;
+
+      dumpUsers(F, "[eraseUselessFunctions] ");
+      for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE;) {
+        auto U = *UI++;
+        if (auto CE = dyn_cast<ConstantExpr>(U)){
+          if (CE->use_empty()) {
+            CE->dropAllReferences();
+            changed = true;
+          }
+        }
+      }
+      if (F->use_empty()) {
+        erase(F);
+        changed = true;
+      }
+    }
+    return changed;
+  }
+
+  void
+  lowerGetBlockInvoke() {
+    if (auto F = M->getFunction(SPIR_INTRINSIC_GET_BLOCK_INVOKE)) {
+      for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE;) {
+        auto CI = dyn_cast<CallInst>(*UI++);
+        assert(CI && "Invalid usage of spir_get_block_invoke");
+        lowerGetBlockInvoke(CI);
+      }
+    }
+  }
+
+  void
+  lowerGetBlockContext() {
+    if (auto F = M->getFunction(SPIR_INTRINSIC_GET_BLOCK_CONTEXT)) {
+      for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE;) {
+        auto CI = dyn_cast<CallInst>(*UI++);
+        assert(CI && "Invalid usage of spir_get_block_context");
+        lowerGetBlockContext(CI);
+      }
+    }
+  }
+  /// Lower calls of spir_block_bind.
+  /// Return true if the Module is changed.
+  bool
+  lowerBlockBind(Function *BlockBindFunc) {
+    bool changed = false;
+    for (auto I = BlockBindFunc->user_begin(), E = BlockBindFunc->user_end();
+        I != E;) {
+      DEBUG(dbgs() << "[lowerBlockBind] " << **I << '\n');
+      // Handle spir_block_bind(bitcast(block_func), context_len,
+      // context_align, context)
+      auto CallBlkBind = cast<CallInst>(*I++);
+      Function *InvF = nullptr;
+      Value *Ctx = nullptr;
+      Value *CtxLen = nullptr;
+      Value *CtxAlign = nullptr;
+      getBlockInvokeFuncAndContext(CallBlkBind, &InvF, &Ctx, &CtxLen,
+          &CtxAlign);
+      for (auto II = CallBlkBind->user_begin(), EE = CallBlkBind->user_end();
+          II != EE;) {
+        auto BlkUser = *II++;
+        SPIRVDBG(dbgs() << "  Block user: " << *BlkUser << '\n');
+        if (auto Ret = dyn_cast<ReturnInst>(BlkUser)) {
+          bool Inlined = false;
+          changed |= lowerReturnBlock(Ret, CallBlkBind, Inlined);
+          if (Inlined)
+            return true;
+        } else if (auto CI = dyn_cast<CallInst>(BlkUser)){
+          auto CallBindF = CI->getCalledFunction();
+          auto Name = CallBindF->getName();
+          std::string DemangledName;
+          if (Name == SPIR_INTRINSIC_GET_BLOCK_INVOKE) {
+            assert(CI->getArgOperand(0) == CallBlkBind);
+            changed |= lowerGetBlockInvoke(CI, cast<Function>(InvF));
+          } else if (Name == SPIR_INTRINSIC_GET_BLOCK_CONTEXT) {
+            assert(CI->getArgOperand(0) == CallBlkBind);
+            // Handle context_ptr = spir_get_block_context(block)
+            lowerGetBlockContext(CI, Ctx);
+            changed = true;
           } else if (oclIsBuiltin(Name, &DemangledName)) {
-            lowerBlockBuiltin(CI, InvF, Ctx, CtxLen, CtxAlign, DemangledName);
-            changed = true;
-          } else
-            llvm_unreachable("Invalid block user");
-        }
-      }
-      erase(CallBlkBind);
-    }
-    changed |= eraseUselessFunctions();
-    return changed;
-  }
-
-  void
-  lowerGetBlockContext(CallInst *CallGetBlkCtx, Value *Ctx = nullptr) {
-    if (!Ctx)
-      getBlockInvokeFuncAndContext(CallGetBlkCtx->getArgOperand(0), nullptr,
-          &Ctx);
-    CallGetBlkCtx->replaceAllUsesWith(Ctx);
-    DEBUG(dbgs() << "  [lowerGetBlockContext] " << *CallGetBlkCtx << " => " <<
-        *Ctx << "\n\n");
-    erase(CallGetBlkCtx);
-  }
-
-  bool
-  lowerGetBlockInvoke(CallInst *CallGetBlkInvoke,
-      Function *InvokeF = nullptr) {
-    bool changed = false;
-    for (auto UI = CallGetBlkInvoke->user_begin(),
-        UE = CallGetBlkInvoke->user_end();
-        UI != UE;) {
-      // Handle block_func_ptr = bitcast(spir_get_block_invoke(block))
-      auto CallInv = cast<Instruction>(*UI++);
-      auto Cast = dyn_cast<BitCastInst>(CallInv);
-      if (Cast)
-        CallInv = dyn_cast<Instruction>(*CallInv->user_begin());
-      DEBUG(dbgs() << "[lowerGetBlockInvoke]  " << *CallInv);
-      // Handle ret = block_func_ptr(context_ptr, args)
-      auto CI = cast<CallInst>(CallInv);
-      auto F = CI->getCalledValue();
-      if (InvokeF == nullptr) {
-        getBlockInvokeFuncAndContext(CallGetBlkInvoke->getArgOperand(0),
-            &InvokeF, nullptr);
-        assert(InvokeF);
-      }
-      assert(F->getType() == InvokeF->getType());
-      CI->replaceUsesOfWith(F, InvokeF);
-      DEBUG(dbgs() << " => " << *CI << "\n\n");
-      erase(Cast);
-      changed = true;
-    }
-    erase(CallGetBlkInvoke);
-    return changed;
-  }
-
-  void
-  lowerBlockBuiltin(CallInst *CI, Function *InvF, Value *Ctx, Value *CtxLen,
-      Value *CtxAlign, const std::string& DemangledName) {
-    mutateCallInstSPIRV (M, CI, [=](CallInst *CI, std::vector<Value *> &Args) {
-      size_t I = 0;
-      size_t E = Args.size();
-      for (; I != E; ++I) {
-        if (isPointerToOpaqueStructType(Args[I]->getType(),
-            SPIR_TYPE_NAME_BLOCK_T)) {
-          break;
-        }
-      }
-      assert (I < E);
-      Args[I] = castToVoidFuncPtr(InvF);
-      if (I + 1 == E) {
-        Args.push_back(Ctx);
-        Args.push_back(CtxLen);
-        Args.push_back(CtxAlign);
-      } else {
-        Args.insert(Args.begin() + I + 1, CtxAlign);
-        Args.insert(Args.begin() + I + 1, CtxLen);
-        Args.insert(Args.begin() + I + 1, Ctx);
-      }
-      if (DemangledName == kOCLBuiltinName::EnqueueKernel) {
-        // Insert event arguments if there are not.
-        if (!isa<IntegerType>(Args[3]->getType())) {
-          Args.insert(Args.begin() + 3, getInt32(M, 0));
-          Args.insert(Args.begin() + 4, getOCLNullClkEventPtr());
-        }
-        if (!isOCLClkEventPtrType(Args[5]->getType()))
-          Args.insert(Args.begin() + 5, getOCLNullClkEventPtr());
-      }
-      return getSPIRVFuncName(OCLSPIRVBuiltinMap::map(DemangledName));
-    });
-  }
-  /// Transform return of a block.
-  /// The function returning a block is inlined since the context cannot be
-  /// passed to another function.
-  /// Returns true of module is changed.
-  bool
-  lowerReturnBlock(ReturnInst *Ret, Value *CallBlkBind, bool &Inlined) {
-    auto F = Ret->getParent()->getParent();
-    auto changed = false;
-    for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE;) {
-      auto U = *UI++;
-      dumpUsers(U);
-      auto Inst = dyn_cast<Instruction>(U);
-      if (Inst && Inst->use_empty()) {
-        erase(Inst);
-        changed = true;
-        continue;
-      }
-      auto CI = dyn_cast<CallInst>(U);
-      if(!CI || CI->getCalledFunction() != F)
-        continue;
-
-      DEBUG(dbgs() << "[lowerReturnBlock] inline " << F->getName() << '\n');
-      auto CG = &getAnalysis<CallGraphWrapperPass>().getCallGraph();
-      auto ACT = &getAnalysis<AssumptionCacheTracker>();
-      auto AA = &getAnalysis<AliasAnalysis>();
-      InlineFunctionInfo IFI(CG, M->getDataLayout(), AA, ACT);
-      InlineFunction(CI, IFI);
-      Inlined = true;
-    }
-    return changed || Inlined;
-  }
-
-  void
-  getBlockInvokeFuncAndContext(Value *Blk, Function **PInvF, Value **PCtx,
-      Value **PCtxLen = nullptr, Value **PCtxAlign = nullptr){
-    Function *InvF = nullptr;
-    Value *Ctx = nullptr;
-    Value *CtxLen = nullptr;
-    Value *CtxAlign = nullptr;
-    if (auto CallBlkBind = dyn_cast<CallInst>(Blk)) {
-      assert(CallBlkBind->getCalledFunction()->getName() ==
-          SPIR_INTRINSIC_BLOCK_BIND && "Invalid block");
-      InvF = dyn_cast<Function>(
-          CallBlkBind->getArgOperand(0)->stripPointerCasts());
-      CtxLen = CallBlkBind->getArgOperand(1);
-      CtxAlign = CallBlkBind->getArgOperand(2);
-      Ctx = CallBlkBind->getArgOperand(3);
-    } else if (auto F = dyn_cast<Function>(Blk->stripPointerCasts())) {
-      InvF = F;
-      Ctx = Constant::getNullValue(IntegerType::getInt8PtrTy(M->getContext()));
-    } else if (auto Load = dyn_cast<LoadInst>(Blk)) {
-      auto Op = Load->getPointerOperand();
-      if (auto GV = dyn_cast<GlobalVariable>(Op)) {
-        if (GV->isConstant()) {
-          InvF = cast<Function>(GV->getInitializer()->stripPointerCasts());
-          Ctx = Constant::getNullValue(IntegerType::getInt8PtrTy(M->getContext()));
-        } else {
-          llvm_unreachable("load non-constant block?");
-        }
-      } else {
-        llvm_unreachable("Loading block from non global?");
-      }
-    } else {
-      llvm_unreachable("Invalid block");
-    }
-    DEBUG(dbgs() << "  Block invocation func: " << InvF->getName() << '\n' <<
-        "  Block context: " << *Ctx << '\n');
-    assert(InvF && Ctx && "Invalid block");
-    if (PInvF)
-      *PInvF = InvF;
-    if (PCtx)
-      *PCtx = Ctx;
-    if (PCtxLen)
-      *PCtxLen = CtxLen;
-    if (PCtxAlign)
-      *PCtxAlign = CtxAlign;
-  }
-  void
-  erase(Instruction *I) {
-    if (!I)
-      return;
-    if (I->use_empty()) {
-      I->dropAllReferences();
-      I->eraseFromParent();
-    }
-    else
-      dumpUsers(I);
-  }
-  void
-  erase(ConstantExpr *I) {
-    if (!I)
-      return;
-    if (I->use_empty()) {
-      I->dropAllReferences();
-      I->destroyConstant();
-    } else
-      dumpUsers(I);
-  }
-  void
-  erase(Function *F) {
-    if (!F)
-      return;
-    if (!F->use_empty()) {
-      dumpUsers(F);
-      return;
-    }
-    F->dropAllReferences();
-    auto &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
-    CG.removeFunctionFromModule(new CallGraphNode(F));
-  }
-
-  llvm::PointerType* getOCLClkEventType() {
-    return getOrCreateOpaquePtrType(M, SPIR_TYPE_NAME_CLK_EVENT_T,
-        SPIRAS_Global);
-  }
-
-  llvm::PointerType* getOCLClkEventPtrType() {
-    return PointerType::get(getOCLClkEventType(), SPIRAS_Generic);
-  }
-
-  bool isOCLClkEventPtrType(Type *T) {
-    if (auto PT = dyn_cast<PointerType>(T))
-      return isPointerToOpaqueStructType(
-        PT->getElementType(), SPIR_TYPE_NAME_CLK_EVENT_T);
-    return false;
-  }
-
-  llvm::Constant* getOCLNullClkEventPtr() {
-    return Constant::getNullValue(getOCLClkEventPtrType());
-  }
-
-  void dumpGetBlockInvokeUsers(StringRef Prompt) {
-    DEBUG(dbgs() << Prompt);
-    dumpUsers(M->getFunction(SPIR_INTRINSIC_GET_BLOCK_INVOKE));
-  }
-};
-
-char SPIRVLowerOCLBlocks::ID = 0;
-}
-
-INITIALIZE_PASS_BEGIN(SPIRVLowerOCLBlocks, "spvblocks",
-    "SPIR-V lower OCL blocks", false, false)
-INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
-INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
-INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
-INITIALIZE_PASS_END(SPIRVLowerOCLBlocks, "spvblocks",
-    "SPIR-V lower OCL blocks", false, false)
-
-ModulePass *llvm::createSPIRVLowerOCLBlocks() {
-  return new SPIRVLowerOCLBlocks();
-}
-
-#endif /* OCLLOWERBLOCKS_H_ */
+            lowerBlockBuiltin(CI, InvF, Ctx, CtxLen, CtxAlign, DemangledName);
+            changed = true;
+          } else
+            llvm_unreachable("Invalid block user");
+        }
+      }
+      erase(CallBlkBind);
+    }
+    changed |= eraseUselessFunctions();
+    return changed;
+  }
+
+  void
+  lowerGetBlockContext(CallInst *CallGetBlkCtx, Value *Ctx = nullptr) {
+    if (!Ctx)
+      getBlockInvokeFuncAndContext(CallGetBlkCtx->getArgOperand(0), nullptr,
+          &Ctx);
+    CallGetBlkCtx->replaceAllUsesWith(Ctx);
+    DEBUG(dbgs() << "  [lowerGetBlockContext] " << *CallGetBlkCtx << " => " <<
+        *Ctx << "\n\n");
+    erase(CallGetBlkCtx);
+  }
+
+  bool
+  lowerGetBlockInvoke(CallInst *CallGetBlkInvoke,
+      Function *InvokeF = nullptr) {
+    bool changed = false;
+    for (auto UI = CallGetBlkInvoke->user_begin(),
+        UE = CallGetBlkInvoke->user_end();
+        UI != UE;) {
+      // Handle block_func_ptr = bitcast(spir_get_block_invoke(block))
+      auto CallInv = cast<Instruction>(*UI++);
+      auto Cast = dyn_cast<BitCastInst>(CallInv);
+      if (Cast)
+        CallInv = dyn_cast<Instruction>(*CallInv->user_begin());
+      DEBUG(dbgs() << "[lowerGetBlockInvoke]  " << *CallInv);
+      // Handle ret = block_func_ptr(context_ptr, args)
+      auto CI = cast<CallInst>(CallInv);
+      auto F = CI->getCalledValue();
+      if (InvokeF == nullptr) {
+        getBlockInvokeFuncAndContext(CallGetBlkInvoke->getArgOperand(0),
+            &InvokeF, nullptr);
+        assert(InvokeF);
+      }
+      assert(F->getType() == InvokeF->getType());
+      CI->replaceUsesOfWith(F, InvokeF);
+      DEBUG(dbgs() << " => " << *CI << "\n\n");
+      erase(Cast);
+      changed = true;
+    }
+    erase(CallGetBlkInvoke);
+    return changed;
+  }
+
+  void
+  lowerBlockBuiltin(CallInst *CI, Function *InvF, Value *Ctx, Value *CtxLen,
+      Value *CtxAlign, const std::string& DemangledName) {
+    mutateCallInstSPIRV (M, CI, [=](CallInst *CI, std::vector<Value *> &Args) {
+      size_t I = 0;
+      size_t E = Args.size();
+      for (; I != E; ++I) {
+        if (isPointerToOpaqueStructType(Args[I]->getType(),
+            SPIR_TYPE_NAME_BLOCK_T)) {
+          break;
+        }
+      }
+      assert (I < E);
+      Args[I] = castToVoidFuncPtr(InvF);
+      if (I + 1 == E) {
+        Args.push_back(Ctx);
+        Args.push_back(CtxLen);
+        Args.push_back(CtxAlign);
+      } else {
+        Args.insert(Args.begin() + I + 1, CtxAlign);
+        Args.insert(Args.begin() + I + 1, CtxLen);
+        Args.insert(Args.begin() + I + 1, Ctx);
+      }
+      if (DemangledName == kOCLBuiltinName::EnqueueKernel) {
+        // Insert event arguments if there are not.
+        if (!isa<IntegerType>(Args[3]->getType())) {
+          Args.insert(Args.begin() + 3, getInt32(M, 0));
+          Args.insert(Args.begin() + 4, getOCLNullClkEventPtr());
+        }
+        if (!isOCLClkEventPtrType(Args[5]->getType()))
+          Args.insert(Args.begin() + 5, getOCLNullClkEventPtr());
+      }
+      return getSPIRVFuncName(OCLSPIRVBuiltinMap::map(DemangledName));
+    });
+  }
+  /// Transform return of a block.
+  /// The function returning a block is inlined since the context cannot be
+  /// passed to another function.
+  /// Returns true of module is changed.
+  bool
+  lowerReturnBlock(ReturnInst *Ret, Value *CallBlkBind, bool &Inlined) {
+    auto F = Ret->getParent()->getParent();
+    auto changed = false;
+    for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE;) {
+      auto U = *UI++;
+      dumpUsers(U);
+      auto Inst = dyn_cast<Instruction>(U);
+      if (Inst && Inst->use_empty()) {
+        erase(Inst);
+        changed = true;
+        continue;
+      }
+      auto CI = dyn_cast<CallInst>(U);
+      if(!CI || CI->getCalledFunction() != F)
+        continue;
+
+      DEBUG(dbgs() << "[lowerReturnBlock] inline " << F->getName() << '\n');
+      auto CG = &getAnalysis<CallGraphWrapperPass>().getCallGraph();
+      auto ACT = &getAnalysis<AssumptionCacheTracker>();
+      //auto AA = &getAnalysis<AliasAnalysis>();
+      //InlineFunctionInfo IFI(CG, M->getDataLayout(), AA, ACT);
+    std::function<AssumptionCache &(Function &)> actf=
+        [&ACT](Function& f) -> AssumptionCache & {return ACT->getAssumptionCache(f);};
+      InlineFunctionInfo IFI(CG,&actf);
+      InlineFunction(CI, IFI);
+      Inlined = true;
+    }
+    return changed || Inlined;
+  }
+
+  void
+  getBlockInvokeFuncAndContext(Value *Blk, Function **PInvF, Value **PCtx,
+      Value **PCtxLen = nullptr, Value **PCtxAlign = nullptr){
+    Function *InvF = nullptr;
+    Value *Ctx = nullptr;
+    Value *CtxLen = nullptr;
+    Value *CtxAlign = nullptr;
+    if (auto CallBlkBind = dyn_cast<CallInst>(Blk)) {
+      assert(CallBlkBind->getCalledFunction()->getName() ==
+          SPIR_INTRINSIC_BLOCK_BIND && "Invalid block");
+      InvF = dyn_cast<Function>(
+          CallBlkBind->getArgOperand(0)->stripPointerCasts());
+      CtxLen = CallBlkBind->getArgOperand(1);
+      CtxAlign = CallBlkBind->getArgOperand(2);
+      Ctx = CallBlkBind->getArgOperand(3);
+    } else if (auto F = dyn_cast<Function>(Blk->stripPointerCasts())) {
+      InvF = F;
+      Ctx = Constant::getNullValue(IntegerType::getInt8PtrTy(M->getContext()));
+    } else if (auto Load = dyn_cast<LoadInst>(Blk)) {
+      auto Op = Load->getPointerOperand();
+      if (auto GV = dyn_cast<GlobalVariable>(Op)) {
+        if (GV->isConstant()) {
+          InvF = cast<Function>(GV->getInitializer()->stripPointerCasts());
+          Ctx = Constant::getNullValue(IntegerType::getInt8PtrTy(M->getContext()));
+        } else {
+          llvm_unreachable("load non-constant block?");
+        }
+      } else {
+        llvm_unreachable("Loading block from non global?");
+      }
+    } else {
+      llvm_unreachable("Invalid block");
+    }
+    DEBUG(dbgs() << "  Block invocation func: " << InvF->getName() << '\n' <<
+        "  Block context: " << *Ctx << '\n');
+    assert(InvF && Ctx && "Invalid block");
+    if (PInvF)
+      *PInvF = InvF;
+    if (PCtx)
+      *PCtx = Ctx;
+    if (PCtxLen)
+      *PCtxLen = CtxLen;
+    if (PCtxAlign)
+      *PCtxAlign = CtxAlign;
+  }
+  void
+  erase(Instruction *I) {
+    if (!I)
+      return;
+    if (I->use_empty()) {
+      I->dropAllReferences();
+      I->eraseFromParent();
+    }
+    else
+      dumpUsers(I);
+  }
+  void
+  erase(ConstantExpr *I) {
+    if (!I)
+      return;
+    if (I->use_empty()) {
+      I->dropAllReferences();
+      I->destroyConstant();
+    } else
+      dumpUsers(I);
+  }
+  void
+  erase(Function *F) {
+    if (!F)
+      return;
+    if (!F->use_empty()) {
+      dumpUsers(F);
+      return;
+    }
+    F->dropAllReferences();
+    auto &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
+    CG.removeFunctionFromModule(new CallGraphNode(F));
+  }
+
+  llvm::PointerType* getOCLClkEventType() {
+    return getOrCreateOpaquePtrType(M, SPIR_TYPE_NAME_CLK_EVENT_T,
+        SPIRAS_Global);
+  }
+
+  llvm::PointerType* getOCLClkEventPtrType() {
+    return PointerType::get(getOCLClkEventType(), SPIRAS_Generic);
+  }
+
+  bool isOCLClkEventPtrType(Type *T) {
+    if (auto PT = dyn_cast<PointerType>(T))
+      return isPointerToOpaqueStructType(
+        PT->getElementType(), SPIR_TYPE_NAME_CLK_EVENT_T);
+    return false;
+  }
+
+  llvm::Constant* getOCLNullClkEventPtr() {
+    return Constant::getNullValue(getOCLClkEventPtrType());
+  }
+
+  void dumpGetBlockInvokeUsers(StringRef Prompt) {
+    DEBUG(dbgs() << Prompt);
+    dumpUsers(M->getFunction(SPIR_INTRINSIC_GET_BLOCK_INVOKE));
+  }
+};
+
+char SPIRVLowerOCLBlocks::ID = 0;
+}
+
+INITIALIZE_PASS_BEGIN(SPIRVLowerOCLBlocks, "spvblocks",
+    "SPIR-V lower OCL blocks", false, false)
+INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
+INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
+initializeAnalysis(Registry);
+INITIALIZE_PASS_END(SPIRVLowerOCLBlocks, "spvblocks",
+    "SPIR-V lower OCL blocks", false, false)
+
+ModulePass *llvm::createSPIRVLowerOCLBlocks() {
+  return new SPIRVLowerOCLBlocks();
+}
+
+#endif /* OCLLOWERBLOCKS_H_ */
diff --git a/SPIRVReader.cpp b/SPIRVReader.cpp
index b4714cf..d29d52b 100644
--- a/SPIRVReader.cpp
+++ b/SPIRVReader.cpp
@@ -1,2511 +1,2513 @@
-//===- SPIRVReader.cpp - Converts SPIR-V to LLVM ----------------*- C++ -*-===//
-//
-//                     The LLVM/SPIR-V Translator
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-// Copyright (c) 2014 Advanced Micro Devices, Inc. All rights reserved.
-//
-// Permission is hereby granted, free of charge, to any person obtaining a
-// copy of this software and associated documentation files (the "Software"),
-// to deal with the Software without restriction, including without limitation
-// the rights to use, copy, modify, merge, publish, distribute, sublicense,
-// and/or sell copies of the Software, and to permit persons to whom the
-// Software is furnished to do so, subject to the following conditions:
-//
-// Redistributions of source code must retain the above copyright notice,
-// this list of conditions and the following disclaimers.
-// Redistributions in binary form must reproduce the above copyright notice,
-// this list of conditions and the following disclaimers in the documentation
-// and/or other materials provided with the distribution.
-// Neither the names of Advanced Micro Devices, Inc., nor the names of its
-// contributors may be used to endorse or promote products derived from this
-// Software without specific prior written permission.
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-// CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
-// THE SOFTWARE.
-//
-//===----------------------------------------------------------------------===//
-/// \file
-///
-/// This file implements conversion of SPIR-V binary to LLVM IR.
-///
-//===----------------------------------------------------------------------===//
-#include "SPIRVUtil.h"
-#include "SPIRVType.h"
-#include "SPIRVValue.h"
-#include "SPIRVModule.h"
-#include "SPIRVFunction.h"
-#include "SPIRVBasicBlock.h"
-#include "SPIRVInstruction.h"
-#include "SPIRVExtInst.h"
-#include "SPIRVInternal.h"
-#include "SPIRVMDBuilder.h"
-#include "OCLUtil.h"
-
-#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/StringSwitch.h"
-#include "llvm/IR/Constants.h"
-#include "llvm/IR/DerivedTypes.h"
-#include "llvm/IR/DIBuilder.h"
-#include "llvm/IR/Instructions.h"
-#include "llvm/IR/Metadata.h"
-#include "llvm/IR/Module.h"
-#include "llvm/IR/Operator.h"
-#include "llvm/IR/Type.h"
-#include "llvm/PassManager.h"
-#include "llvm/Support/Casting.h"
-#include "llvm/Support/Debug.h"
-#include "llvm/Support/Dwarf.h"
-#include "llvm/Support/FileSystem.h"
-#include "llvm/Support/raw_ostream.h"
-#include "llvm/Support/CommandLine.h"
-
-#include <algorithm>
-#include <cstdlib>
-#include <functional>
-#include <fstream>
-#include <iostream>
-#include <iterator>
-#include <map>
-#include <set>
-#include <sstream>
-#include <string>
-
-#define DEBUG_TYPE "spirv"
-
-using namespace std;
-using namespace llvm;
-using namespace SPIRV;
-using namespace OCLUtil;
-
-namespace SPIRV{
-
-cl::opt<bool> SPIRVEnableStepExpansion("spirv-expand-step", cl::init(true),
-  cl::desc("Enable expansion of OpenCL step and smoothstep function"));
-
-cl::opt<bool> SPIRVGenKernelArgNameMD("spirv-gen-kernel-arg-name-md",
-    cl::init(false), cl::desc("Enable generating OpenCL kernel argument name "
-    "metadata"));
-
-cl::opt<bool> SPIRVGenImgTypeAccQualPostfix("spirv-gen-image-type-acc-postfix",
-    cl::init(false), cl::desc("Enable generating access qualifier postfix"
-        " in OpenCL image type names"));
-
-// Prefix for placeholder global variable name.
-const char* kPlaceholderPrefix = "placeholder.";
-
-// Save the translated LLVM before validation for debugging purpose.
-static bool DbgSaveTmpLLVM = true;
-static const char *DbgTmpLLVMFileName = "_tmp_llvmbil.ll";
-
-typedef std::pair < unsigned, AttributeSet > AttributeWithIndex;
-
-static bool
-isOpenCLKernel(SPIRVFunction *BF) {
-  return BF->getModule()->isEntryPoint(ExecutionModelKernel, BF->getId());
-}
-
-static void
-dumpLLVM(Module *M, const std::string &FName) {
-  std::error_code EC;
-  raw_fd_ostream FS(FName, EC, sys::fs::F_None);
-  if (EC) {
-    FS << *M;
-    FS.close();
-  }
-}
-
-static MDNode*
-getMDNodeStringIntVec(LLVMContext *Context, const std::string& Str,
-    const std::vector<SPIRVWord>& IntVals) {
-  std::vector<Metadata*> ValueVec;
-  ValueVec.push_back(MDString::get(*Context, Str));
-  for (auto &I:IntVals)
-    ValueVec.push_back(ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(*Context), I)));
-  return MDNode::get(*Context, ValueVec);
-}
-
-static MDNode*
-getMDTwoInt(LLVMContext *Context, unsigned Int1, unsigned Int2) {
-  std::vector<Metadata*> ValueVec;
-  ValueVec.push_back(ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(*Context), Int1)));
-  ValueVec.push_back(ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(*Context), Int2)));
-  return MDNode::get(*Context, ValueVec);
-}
-
-static MDNode*
-getMDString(LLVMContext *Context, const std::string& Str) {
-  std::vector<Metadata*> ValueVec;
-  if (!Str.empty())
-    ValueVec.push_back(MDString::get(*Context, Str));
-  return MDNode::get(*Context, ValueVec);
-}
-
-static void
-addOCLVersionMetadata(LLVMContext *Context, Module *M,
-    const std::string &MDName, unsigned Major, unsigned Minor) {
-  NamedMDNode *NamedMD = M->getOrInsertNamedMetadata(MDName);
-  NamedMD->addOperand(getMDTwoInt(Context, Major, Minor));
-}
-
-static void
-addNamedMetadataStringSet(LLVMContext *Context, Module *M,
-    const std::string &MDName, const std::set<std::string> &StrSet) {
-  NamedMDNode *NamedMD = M->getOrInsertNamedMetadata(MDName);
-  std::vector<Metadata*> ValueVec;
-  for (auto &&Str : StrSet) {
-    ValueVec.push_back(MDString::get(*Context, Str));
-  }
-  NamedMD->addOperand(MDNode::get(*Context, ValueVec));
-}
-
-static void
-addOCLKernelArgumentMetadata(LLVMContext *Context,
-  std::vector<llvm::Metadata*> &KernelMD, const std::string &MDName,
-    SPIRVFunction *BF, std::function<Metadata *(SPIRVFunctionParameter *)>Func){
-  std::vector<Metadata*> ValueVec;
-    ValueVec.push_back(MDString::get(*Context, MDName));
-  BF->foreachArgument([&](SPIRVFunctionParameter *Arg) {
-    ValueVec.push_back(Func(Arg));
-  });
-  KernelMD.push_back(MDNode::get(*Context, ValueVec));
-}
-
-class SPIRVToLLVMDbgTran {
-public:
-  SPIRVToLLVMDbgTran(SPIRVModule *TBM, Module *TM)
-  :BM(TBM), M(TM), SpDbg(BM), Builder(*M){
-    Enable = BM->hasDebugInfo();
-  }
-
-  void createCompileUnit() {
-    if (!Enable)
-      return;
-    auto File = SpDbg.getEntryPointFileStr(ExecutionModelKernel, 0);
-    std::string BaseName;
-    std::string Path;
-    splitFileName(File, BaseName, Path);
-    Builder.createCompileUnit(dwarf::DW_LANG_C99,
-      BaseName, Path, "spirv", false, "", 0, "", DIBuilder::LineTablesOnly);
-  }
-
-  void addDbgInfoVersion() {
-    if (!Enable)
-      return;
-    M->addModuleFlag(Module::Warning, "Dwarf Version",
-        dwarf::DWARF_VERSION);
-    M->addModuleFlag(Module::Warning, "Debug Info Version",
-        DEBUG_METADATA_VERSION);
-  }
-
-  DIFile getDIFile(const std::string &FileName){
-    return getOrInsert(FileMap, FileName, [=](){
-      std::string BaseName;
-      std::string Path;
-      splitFileName(FileName, BaseName, Path);
-      if (!BaseName.empty())
-        return Builder.createFile(BaseName, Path);
-      else
-        return DIFile();
-    });
-  }
-
-  DISubprogram getDISubprogram(SPIRVFunction *SF, Function *F){
-    return getOrInsert(FuncMap, F, [=](){
-      auto DF = getDIFile(SpDbg.getFunctionFileStr(SF));
-      auto FN = F->getName();
-      auto LN = SpDbg.getFunctionLineNo(SF);
-      Metadata *Args[] = {DIType()};
-      return Builder.createFunction(DF, FN, FN, DF, LN,
-        Builder.createSubroutineType(DF, Builder.getOrCreateTypeArray(Args)),
-        Function::isInternalLinkage(F->getLinkage()),
-        true, LN, 0, 0, NULL, NULL, NULL);
-    });
-  }
-
-  void transDbgInfo(SPIRVValue *SV, Value *V) {
-    if (!Enable || !SV->hasLine())
-      return;
-    if (auto I = dyn_cast<Instruction>(V)) {
-      assert(SV->isInst() && "Invalid instruction");
-      auto SI = static_cast<SPIRVInstruction *>(SV);
-      assert(SI->getParent() &&
-             SI->getParent()->getParent() &&
-             "Invalid instruction");
-      auto Line = SV->getLine();
-      I->setDebugLoc(DebugLoc::get(Line->getLine(), Line->getColumn(),
-          getDISubprogram(SI->getParent()->getParent(),
-              I->getParent()->getParent())));
-    }
-  }
-
-  void finalize() {
-    if (!Enable)
-      return;
-    Builder.finalize();
-  }
-
-private:
-  SPIRVModule *BM;
-  Module *M;
-  SPIRVDbgInfo SpDbg;
-  DIBuilder Builder;
-  bool Enable;
-  std::unordered_map<std::string, DIFile> FileMap;
-  std::unordered_map<Function *, DISubprogram> FuncMap;
-
-  void splitFileName(const std::string &FileName,
-      std::string &BaseName,
-      std::string &Path) {
-    auto Loc = FileName.find_last_of("/\\");
-    if (Loc != std::string::npos) {
-      BaseName = FileName.substr(Loc + 1);
-      Path = FileName.substr(0, Loc);
-    } else {
-      BaseName = FileName;
-      Path = ".";
-    }
-  }
-};
-
-class SPIRVToLLVM {
-public:
-  SPIRVToLLVM(Module *LLVMModule, SPIRVModule *TheSPIRVModule)
-    :M(LLVMModule), BM(TheSPIRVModule), DbgTran(BM, M){
-    assert(M);
-    Context = &M->getContext();
-  }
-
-  std::string getOCLBuiltinName(SPIRVInstruction* BI);
-  std::string getOCLConvertBuiltinName(SPIRVInstruction *BI);
-  std::string getOCLGenericCastToPtrName(SPIRVInstruction *BI);
-
-  Type *transType(SPIRVType *BT);
-  std::string transTypeToOCLTypeName(SPIRVType *BT, bool IsSigned = true);
-  std::vector<Type *> transTypeVector(const std::vector<SPIRVType *>&);
-  bool translate();
-  bool transAddressingModel();
-
-  Value *transValue(SPIRVValue *, Function *F, BasicBlock *,
-      bool CreatePlaceHolder = true);
-  Value *transValueWithoutDecoration(SPIRVValue *, Function *F, BasicBlock *,
-      bool CreatePlaceHolder = true);
-  bool transDecoration(SPIRVValue *, Value *);
-  bool transAlign(SPIRVValue *, Value *);
-  Instruction *transOCLBuiltinFromExtInst(SPIRVExtInst *BC, BasicBlock *BB);
-  std::vector<Value *> transValue(const std::vector<SPIRVValue *>&, Function *F,
-      BasicBlock *);
-  Function *transFunction(SPIRVFunction *F);
-  bool transFPContractMetadata();
-  bool transKernelMetadata();
-  bool transSourceLanguage();
-  bool transSourceExtension();
-  void transGeneratorMD();
-  Value *transConvertInst(SPIRVValue* BV, Function* F, BasicBlock* BB);
-  Instruction *transBuiltinFromInst(const std::string& FuncName,
-      SPIRVInstruction* BI, BasicBlock* BB);
-  Instruction *transOCLBuiltinFromInst(SPIRVInstruction *BI, BasicBlock *BB);
-  Instruction *transSPIRVBuiltinFromInst(SPIRVInstruction *BI, BasicBlock *BB);
-  Instruction *transOCLBarrierFence(SPIRVInstruction* BI, BasicBlock *BB);
-  void transOCLVectorLoadStore(std::string& UnmangledName,
-      std::vector<SPIRVWord> &BArgs);
-
-  /// Post-process translated LLVM module for OpenCL.
-  bool postProcessOCL();
-
-  /// \brief Post-process OpenCL builtin functions returning struct type.
-  ///
-  /// Some OpenCL builtin functions are translated to SPIR-V instructions with
-  /// struct type result, e.g. NDRange creation functions. Such functions
-  /// need to be post-processed to return the struct through sret argument.
-  bool postProcessOCLBuiltinReturnStruct(Function *F);
-
-  /// \brief Post-process OpenCL builtin functions having block argument.
-  ///
-  /// These functions are translated to functions with function pointer type
-  /// argument first, then post-processed to have block argument.
-  bool postProcessOCLBuiltinWithFuncPointer(Function *F,
-      Function::arg_iterator I);
-
-  /// \brief Post-process OpenCL builtin functions having array argument.
-  ///
-  /// These functions are translated to functions with array type argument
-  /// first, then post-processed to have pointer arguments.
-  bool postProcessOCLBuiltinWithArrayArguments(Function *F,
-      const std::string &DemangledName);
-
-  /// \brief Post-process OpImageSampleExplicitLod.
-  ///   sampled_image = __spirv_SampledImage__(image, sampler);
-  ///   return __spirv_ImageSampleExplicitLod__(sampled_image, image_operands,
-  ///                                           ...);
-  /// =>
-  ///   read_image(image, sampler, ...)
-  /// \return transformed call instruction.
-  Instruction *postProcessOCLReadImage(SPIRVInstruction *BI, CallInst *CI,
-      const std::string &DemangledName);
-
-  /// \brief Post-process OpImageWrite.
-  ///   return write_image(image, coord, color, image_operands, ...);
-  /// =>
-  ///   write_image(image, coord, ..., color)
-  /// \return transformed call instruction.
-  CallInst *postProcessOCLWriteImage(SPIRVInstruction *BI, CallInst *CI,
-      const std::string &DemangledName);
-
-  /// \brief Post-process OpBuildNDRange.
-  ///   OpBuildNDRange GlobalWorkSize, LocalWorkSize, GlobalWorkOffset
-  /// =>
-  ///   call ndrange_XD(GlobalWorkOffset, GlobalWorkSize, LocalWorkSize)
-  /// \return transformed call instruction.
-  CallInst *postProcessOCLBuildNDRange(SPIRVInstruction *BI, CallInst *CI,
-      const std::string &DemangledName);
-
-  /// \brief Expand OCL builtin functions with scalar argument, e.g.
-  /// step, smoothstep.
-  /// gentype func (fp edge, gentype x)
-  /// =>
-  /// gentype func (gentype edge, gentype x)
-  /// \return transformed call instruction.
-  CallInst *expandOCLBuiltinWithScalarArg(CallInst* CI,
-      const std::string &FuncName);
-
-  /// \brief Post-process OpGroupAll and OpGroupAny instructions translation.
-  /// i1 func (<n x i1> arg)
-  /// =>
-  /// i32 func (<n x i32> arg)
-  /// \return transformed call instruction.
-  Instruction *postProcessGroupAllAny(CallInst *CI,
-                                      const std::string &DemangledName);
-
-  typedef DenseMap<SPIRVType *, Type *> SPIRVToLLVMTypeMap;
-  typedef DenseMap<SPIRVValue *, Value *> SPIRVToLLVMValueMap;
-  typedef DenseMap<SPIRVFunction *, Function *> SPIRVToLLVMFunctionMap;
-  typedef DenseMap<GlobalVariable *, SPIRVBuiltinVariableKind> BuiltinVarMap;
-
-  // A SPIRV value may be translated to a load instruction of a placeholder
-  // global variable. This map records load instruction of these placeholders
-  // which are supposed to be replaced by the real values later.
-  typedef std::map<SPIRVValue *, LoadInst*> SPIRVToLLVMPlaceholderMap;
-private:
-  Module *M;
-  BuiltinVarMap BuiltinGVMap;
-  LLVMContext *Context;
-  SPIRVModule *BM;
-  SPIRVToLLVMTypeMap TypeMap;
-  SPIRVToLLVMValueMap ValueMap;
-  SPIRVToLLVMFunctionMap FuncMap;
-  SPIRVToLLVMPlaceholderMap PlaceholderMap;
-  SPIRVToLLVMDbgTran DbgTran;
-
-  Type *mapType(SPIRVType *BT, Type *T) {
-    SPIRVDBG(dbgs() << *T << '\n';)
-    TypeMap[BT] = T;
-    return T;
-  }
-
-  // If a value is mapped twice, the existing mapped value is a placeholder,
-  // which must be a load instruction of a global variable whose name starts
-  // with kPlaceholderPrefix.
-  Value *mapValue(SPIRVValue *BV, Value *V) {
-    auto Loc = ValueMap.find(BV);
-    if (Loc != ValueMap.end()) {
-      if (Loc->second == V)
-        return V;
-      auto LD = dyn_cast<LoadInst>(Loc->second);
-      auto Placeholder = dyn_cast<GlobalVariable>(LD->getPointerOperand());
-      assert (LD && Placeholder &&
-          Placeholder->getName().startswith(kPlaceholderPrefix) &&
-          "A value is translated twice");
-      // Replaces placeholders for PHI nodes
-      LD->replaceAllUsesWith(V);
-      LD->dropAllReferences();
-      LD->removeFromParent();
-      Placeholder->dropAllReferences();
-      Placeholder->removeFromParent();
-    }
-    ValueMap[BV] = V;
-    return V;
-  }
-
-  bool isSPIRVBuiltinVariable(GlobalVariable *GV,
-      SPIRVBuiltinVariableKind *Kind = nullptr) {
-    auto Loc = BuiltinGVMap.find(GV);
-    if (Loc == BuiltinGVMap.end())
-      return false;
-    if (Kind)
-      *Kind = Loc->second;
-    return true;
-  }
-  // OpenCL function always has NoUnwound attribute.
-  // Change this if it is no longer true.
-  bool isFuncNoUnwind() const { return true;}
-  bool isSPIRVCmpInstTransToLLVMInst(SPIRVInstruction *BI) const;
-  bool transOCLBuiltinsFromVariables();
-  bool transOCLBuiltinFromVariable(GlobalVariable *GV,
-      SPIRVBuiltinVariableKind Kind);
-  MDString *transOCLKernelArgTypeName(SPIRVFunctionParameter *);
-
-  Value *mapFunction(SPIRVFunction *BF, Function *F) {
-    SPIRVDBG(spvdbgs() << "[mapFunction] " << *BF << " -> ";
-      dbgs() << *F << '\n';)
-    FuncMap[BF] = F;
-    return F;
-  }
-
-  Value *getTranslatedValue(SPIRVValue *BV);
-  Type *getTranslatedType(SPIRVType *BT);
-
-  SPIRVErrorLog &getErrorLog() {
-    return BM->getErrorLog();
-  }
-
-  void setCallingConv(CallInst *Call) {
-    Function *F = Call->getCalledFunction();
+//===- SPIRVReader.cpp - Converts SPIR-V to LLVM ----------------*- C++ -*-===//
+//
+//                     The LLVM/SPIR-V Translator
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+// Copyright (c) 2014 Advanced Micro Devices, Inc. All rights reserved.
+//
+// Permission is hereby granted, free of charge, to any person obtaining a
+// copy of this software and associated documentation files (the "Software"),
+// to deal with the Software without restriction, including without limitation
+// the rights to use, copy, modify, merge, publish, distribute, sublicense,
+// and/or sell copies of the Software, and to permit persons to whom the
+// Software is furnished to do so, subject to the following conditions:
+//
+// Redistributions of source code must retain the above copyright notice,
+// this list of conditions and the following disclaimers.
+// Redistributions in binary form must reproduce the above copyright notice,
+// this list of conditions and the following disclaimers in the documentation
+// and/or other materials provided with the distribution.
+// Neither the names of Advanced Micro Devices, Inc., nor the names of its
+// contributors may be used to endorse or promote products derived from this
+// Software without specific prior written permission.
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+// CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+// THE SOFTWARE.
+//
+//===----------------------------------------------------------------------===//
+/// \file
+///
+/// This file implements conversion of SPIR-V binary to LLVM IR.
+///
+//===----------------------------------------------------------------------===//
+#include "SPIRVUtil.h"
+#include "SPIRVType.h"
+#include "SPIRVValue.h"
+#include "SPIRVModule.h"
+#include "SPIRVFunction.h"
+#include "SPIRVBasicBlock.h"
+#include "SPIRVInstruction.h"
+#include "SPIRVExtInst.h"
+#include "SPIRVInternal.h"
+#include "SPIRVMDBuilder.h"
+#include "OCLUtil.h"
+
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/StringSwitch.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/DIBuilder.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/Metadata.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/Operator.h"
+#include "llvm/IR/Type.h"
+#include "llvm/IR/LegacyPassManager.h"
+#include "llvm/Support/Casting.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/Dwarf.h"
+#include "llvm/Support/FileSystem.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Support/CommandLine.h"
+
+#include <algorithm>
+#include <cstdlib>
+#include <functional>
+#include <fstream>
+#include <iostream>
+#include <iterator>
+#include <map>
+#include <set>
+#include <sstream>
+#include <string>
+
+#define DEBUG_TYPE "spirv"
+
+using namespace std;
+using namespace llvm;
+using namespace SPIRV;
+using namespace OCLUtil;
+
+namespace SPIRV{
+
+cl::opt<bool> SPIRVEnableStepExpansion("spirv-expand-step", cl::init(true),
+  cl::desc("Enable expansion of OpenCL step and smoothstep function"));
+
+cl::opt<bool> SPIRVGenKernelArgNameMD("spirv-gen-kernel-arg-name-md",
+    cl::init(false), cl::desc("Enable generating OpenCL kernel argument name "
+    "metadata"));
+
+cl::opt<bool> SPIRVGenImgTypeAccQualPostfix("spirv-gen-image-type-acc-postfix",
+    cl::init(false), cl::desc("Enable generating access qualifier postfix"
+        " in OpenCL image type names"));
+
+// Prefix for placeholder global variable name.
+const char* kPlaceholderPrefix = "placeholder.";
+
+// Save the translated LLVM before validation for debugging purpose.
+static bool DbgSaveTmpLLVM = true;
+static const char *DbgTmpLLVMFileName = "_tmp_llvmbil.ll";
+
+typedef std::pair < unsigned, AttributeSet > AttributeWithIndex;
+
+static bool
+isOpenCLKernel(SPIRVFunction *BF) {
+  return BF->getModule()->isEntryPoint(ExecutionModelKernel, BF->getId());
+}
+
+static void
+dumpLLVM(Module *M, const std::string &FName) {
+  std::error_code EC;
+  raw_fd_ostream FS(FName, EC, sys::fs::F_None);
+  if (EC) {
+    FS << *M;
+    FS.close();
+  }
+}
+
+static MDNode*
+getMDNodeStringIntVec(LLVMContext *Context, const std::string& Str,
+    const std::vector<SPIRVWord>& IntVals) {
+  std::vector<Metadata*> ValueVec;
+  ValueVec.push_back(MDString::get(*Context, Str));
+  for (auto &I:IntVals)
+    ValueVec.push_back(ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(*Context), I)));
+  return MDNode::get(*Context, ValueVec);
+}
+
+static MDNode*
+getMDTwoInt(LLVMContext *Context, unsigned Int1, unsigned Int2) {
+  std::vector<Metadata*> ValueVec;
+  ValueVec.push_back(ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(*Context), Int1)));
+  ValueVec.push_back(ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(*Context), Int2)));
+  return MDNode::get(*Context, ValueVec);
+}
+
+static MDNode*
+getMDString(LLVMContext *Context, const std::string& Str) {
+  std::vector<Metadata*> ValueVec;
+  if (!Str.empty())
+    ValueVec.push_back(MDString::get(*Context, Str));
+  return MDNode::get(*Context, ValueVec);
+}
+
+static void
+addOCLVersionMetadata(LLVMContext *Context, Module *M,
+    const std::string &MDName, unsigned Major, unsigned Minor) {
+  NamedMDNode *NamedMD = M->getOrInsertNamedMetadata(MDName);
+  NamedMD->addOperand(getMDTwoInt(Context, Major, Minor));
+}
+
+static void
+addNamedMetadataStringSet(LLVMContext *Context, Module *M,
+    const std::string &MDName, const std::set<std::string> &StrSet) {
+  NamedMDNode *NamedMD = M->getOrInsertNamedMetadata(MDName);
+  std::vector<Metadata*> ValueVec;
+  for (auto &&Str : StrSet) {
+    ValueVec.push_back(MDString::get(*Context, Str));
+  }
+  NamedMD->addOperand(MDNode::get(*Context, ValueVec));
+}
+
+static void
+addOCLKernelArgumentMetadata(LLVMContext *Context,
+  std::vector<llvm::Metadata*> &KernelMD, const std::string &MDName,
+    SPIRVFunction *BF, std::function<Metadata *(SPIRVFunctionParameter *)>Func){
+  std::vector<Metadata*> ValueVec;
+    ValueVec.push_back(MDString::get(*Context, MDName));
+  BF->foreachArgument([&](SPIRVFunctionParameter *Arg) {
+    ValueVec.push_back(Func(Arg));
+  });
+  KernelMD.push_back(MDNode::get(*Context, ValueVec));
+}
+
+class SPIRVToLLVMDbgTran {
+public:
+  SPIRVToLLVMDbgTran(SPIRVModule *TBM, Module *TM)
+  :BM(TBM), M(TM), SpDbg(BM), Builder(*M){
+    Enable = BM->hasDebugInfo();
+  }
+
+  void createCompileUnit() {
+    if (!Enable)
+      return;
+    auto File = SpDbg.getEntryPointFileStr(ExecutionModelKernel, 0);
+    //std::string BaseName;
+    //std::string Path;
+    //splitFileName(File, BaseName, Path);
+    Builder.createCompileUnit(dwarf::DW_LANG_C99,
+      getDIFile(File), "spirv", false, "", 0, "", DICompileUnit::DebugEmissionKind::LineTablesOnly);
+  }
+
+  void addDbgInfoVersion() {
+    if (!Enable)
+      return;
+    M->addModuleFlag(Module::Warning, "Dwarf Version",
+        dwarf::DWARF_VERSION);
+    M->addModuleFlag(Module::Warning, "Debug Info Version",
+        DEBUG_METADATA_VERSION);
+  }
+
+  DIFile *getDIFile(const std::string &FileName){
+    return getOrInsert(FileMap, FileName, [=](){
+      std::string BaseName;
+      std::string Path;
+      splitFileName(FileName, BaseName, Path);
+      if (!BaseName.empty())
+        return Builder.createFile(BaseName, Path);
+      else
+        return (DIFile *)nullptr;
+    });
+  }
+
+  DISubprogram* getDISubprogram(SPIRVFunction *SF, Function *F){
+    return getOrInsert(FuncMap, F, [=](){
+      auto DF = getDIFile(SpDbg.getFunctionFileStr(SF));
+      auto FN = F->getName();
+      auto LN = SpDbg.getFunctionLineNo(SF);
+      Metadata *Args[] = {nullptr};
+      return Builder.createFunction(DF, FN, FN, DF, LN,
+        Builder.createSubroutineType(Builder.getOrCreateTypeArray({Args})),
+        Function::isInternalLinkage(F->getLinkage()),
+        true, LN);
+    });
+  }
+
+  void transDbgInfo(SPIRVValue *SV, Value *V) {
+    if (!Enable || !SV->hasLine())
+      return;
+    if (auto I = dyn_cast<Instruction>(V)) {
+      assert(SV->isInst() && "Invalid instruction");
+      auto SI = static_cast<SPIRVInstruction *>(SV);
+      assert(SI->getParent() &&
+             SI->getParent()->getParent() &&
+             "Invalid instruction");
+      auto Line = SV->getLine();
+      I->setDebugLoc(DebugLoc::get(Line->getLine(), Line->getColumn(),
+          getDISubprogram(SI->getParent()->getParent(),
+              I->getParent()->getParent())));
+    }
+  }
+
+  void finalize() {
+    if (!Enable)
+      return;
+    Builder.finalize();
+  }
+
+private:
+  SPIRVModule *BM;
+  Module *M;
+  SPIRVDbgInfo SpDbg;
+  DIBuilder Builder;
+  bool Enable;
+  std::unordered_map<std::string, DIFile*> FileMap;
+  std::unordered_map<Function *, DISubprogram*> FuncMap;
+
+  void splitFileName(const std::string &FileName,
+      std::string &BaseName,
+      std::string &Path) {
+    auto Loc = FileName.find_last_of("/\\");
+    if (Loc != std::string::npos) {
+      BaseName = FileName.substr(Loc + 1);
+      Path = FileName.substr(0, Loc);
+    } else {
+      BaseName = FileName;
+      Path = ".";
+    }
+  }
+};
+
+class SPIRVToLLVM {
+public:
+  SPIRVToLLVM(Module *LLVMModule, SPIRVModule *TheSPIRVModule)
+    :M(LLVMModule), BM(TheSPIRVModule), DbgTran(BM, M){
+    assert(M);
+    Context = &M->getContext();
+  }
+
+  std::string getOCLBuiltinName(SPIRVInstruction* BI);
+  std::string getOCLConvertBuiltinName(SPIRVInstruction *BI);
+  std::string getOCLGenericCastToPtrName(SPIRVInstruction *BI);
+
+  Type *transType(SPIRVType *BT);
+  std::string transTypeToOCLTypeName(SPIRVType *BT, bool IsSigned = true);
+  std::vector<Type *> transTypeVector(const std::vector<SPIRVType *>&);
+  bool translate();
+  bool transAddressingModel();
+
+  Value *transValue(SPIRVValue *, Function *F, BasicBlock *,
+      bool CreatePlaceHolder = true);
+  Value *transValueWithoutDecoration(SPIRVValue *, Function *F, BasicBlock *,
+      bool CreatePlaceHolder = true);
+  bool transDecoration(SPIRVValue *, Value *);
+  bool transAlign(SPIRVValue *, Value *);
+  Instruction *transOCLBuiltinFromExtInst(SPIRVExtInst *BC, BasicBlock *BB);
+  std::vector<Value *> transValue(const std::vector<SPIRVValue *>&, Function *F,
+      BasicBlock *);
+  Function *transFunction(SPIRVFunction *F);
+  bool transFPContractMetadata();
+  bool transKernelMetadata();
+  bool transSourceLanguage();
+  bool transSourceExtension();
+  void transGeneratorMD();
+  Value *transConvertInst(SPIRVValue* BV, Function* F, BasicBlock* BB);
+  Instruction *transBuiltinFromInst(const std::string& FuncName,
+      SPIRVInstruction* BI, BasicBlock* BB);
+  Instruction *transOCLBuiltinFromInst(SPIRVInstruction *BI, BasicBlock *BB);
+  Instruction *transSPIRVBuiltinFromInst(SPIRVInstruction *BI, BasicBlock *BB);
+  Instruction *transOCLBarrierFence(SPIRVInstruction* BI, BasicBlock *BB);
+  void transOCLVectorLoadStore(std::string& UnmangledName,
+      std::vector<SPIRVWord> &BArgs);
+
+  /// Post-process translated LLVM module for OpenCL.
+  bool postProcessOCL();
+
+  /// \brief Post-process OpenCL builtin functions returning struct type.
+  ///
+  /// Some OpenCL builtin functions are translated to SPIR-V instructions with
+  /// struct type result, e.g. NDRange creation functions. Such functions
+  /// need to be post-processed to return the struct through sret argument.
+  bool postProcessOCLBuiltinReturnStruct(Function *F);
+
+  /// \brief Post-process OpenCL builtin functions having block argument.
+  ///
+  /// These functions are translated to functions with function pointer type
+  /// argument first, then post-processed to have block argument.
+  bool postProcessOCLBuiltinWithFuncPointer(Function *F,
+      Function::arg_iterator I);
+
+  /// \brief Post-process OpenCL builtin functions having array argument.
+  ///
+  /// These functions are translated to functions with array type argument
+  /// first, then post-processed to have pointer arguments.
+  bool postProcessOCLBuiltinWithArrayArguments(Function *F,
+      const std::string &DemangledName);
+
+  /// \brief Post-process OpImageSampleExplicitLod.
+  ///   sampled_image = __spirv_SampledImage__(image, sampler);
+  ///   return __spirv_ImageSampleExplicitLod__(sampled_image, image_operands,
+  ///                                           ...);
+  /// =>
+  ///   read_image(image, sampler, ...)
+  /// \return transformed call instruction.
+  Instruction *postProcessOCLReadImage(SPIRVInstruction *BI, CallInst *CI,
+      const std::string &DemangledName);
+
+  /// \brief Post-process OpImageWrite.
+  ///   return write_image(image, coord, color, image_operands, ...);
+  /// =>
+  ///   write_image(image, coord, ..., color)
+  /// \return transformed call instruction.
+  CallInst *postProcessOCLWriteImage(SPIRVInstruction *BI, CallInst *CI,
+      const std::string &DemangledName);
+
+  /// \brief Post-process OpBuildNDRange.
+  ///   OpBuildNDRange GlobalWorkSize, LocalWorkSize, GlobalWorkOffset
+  /// =>
+  ///   call ndrange_XD(GlobalWorkOffset, GlobalWorkSize, LocalWorkSize)
+  /// \return transformed call instruction.
+  CallInst *postProcessOCLBuildNDRange(SPIRVInstruction *BI, CallInst *CI,
+      const std::string &DemangledName);
+
+  /// \brief Expand OCL builtin functions with scalar argument, e.g.
+  /// step, smoothstep.
+  /// gentype func (fp edge, gentype x)
+  /// =>
+  /// gentype func (gentype edge, gentype x)
+  /// \return transformed call instruction.
+  CallInst *expandOCLBuiltinWithScalarArg(CallInst* CI,
+      const std::string &FuncName);
+
+  /// \brief Post-process OpGroupAll and OpGroupAny instructions translation.
+  /// i1 func (<n x i1> arg)
+  /// =>
+  /// i32 func (<n x i32> arg)
+  /// \return transformed call instruction.
+  Instruction *postProcessGroupAllAny(CallInst *CI,
+                                      const std::string &DemangledName);
+
+  typedef DenseMap<SPIRVType *, Type *> SPIRVToLLVMTypeMap;
+  typedef DenseMap<SPIRVValue *, Value *> SPIRVToLLVMValueMap;
+  typedef DenseMap<SPIRVFunction *, Function *> SPIRVToLLVMFunctionMap;
+  typedef DenseMap<GlobalVariable *, SPIRVBuiltinVariableKind> BuiltinVarMap;
+
+  // A SPIRV value may be translated to a load instruction of a placeholder
+  // global variable. This map records load instruction of these placeholders
+  // which are supposed to be replaced by the real values later.
+  typedef std::map<SPIRVValue *, LoadInst*> SPIRVToLLVMPlaceholderMap;
+private:
+  Module *M;
+  BuiltinVarMap BuiltinGVMap;
+  LLVMContext *Context;
+  SPIRVModule *BM;
+  SPIRVToLLVMTypeMap TypeMap;
+  SPIRVToLLVMValueMap ValueMap;
+  SPIRVToLLVMFunctionMap FuncMap;
+  SPIRVToLLVMPlaceholderMap PlaceholderMap;
+  SPIRVToLLVMDbgTran DbgTran;
+
+  Type *mapType(SPIRVType *BT, Type *T) {
+    SPIRVDBG(dbgs() << *T << '\n';)
+    TypeMap[BT] = T;
+    return T;
+  }
+
+  // If a value is mapped twice, the existing mapped value is a placeholder,
+  // which must be a load instruction of a global variable whose name starts
+  // with kPlaceholderPrefix.
+  Value *mapValue(SPIRVValue *BV, Value *V) {
+    auto Loc = ValueMap.find(BV);
+    if (Loc != ValueMap.end()) {
+      if (Loc->second == V)
+        return V;
+      auto LD = dyn_cast<LoadInst>(Loc->second);
+      auto Placeholder = dyn_cast<GlobalVariable>(LD->getPointerOperand());
+      assert (LD && Placeholder &&
+          Placeholder->getName().startswith(kPlaceholderPrefix) &&
+          "A value is translated twice");
+      // Replaces placeholders for PHI nodes
+      LD->replaceAllUsesWith(V);
+      LD->dropAllReferences();
+      LD->removeFromParent();
+      Placeholder->dropAllReferences();
+      Placeholder->removeFromParent();
+    }
+    ValueMap[BV] = V;
+    return V;
+  }
+
+  bool isSPIRVBuiltinVariable(GlobalVariable *GV,
+      SPIRVBuiltinVariableKind *Kind = nullptr) {
+    auto Loc = BuiltinGVMap.find(GV);
+    if (Loc == BuiltinGVMap.end())
+      return false;
+    if (Kind)
+      *Kind = Loc->second;
+    return true;
+  }
+  // OpenCL function always has NoUnwound attribute.
+  // Change this if it is no longer true.
+  bool isFuncNoUnwind() const { return true;}
+  bool isSPIRVCmpInstTransToLLVMInst(SPIRVInstruction *BI) const;
+  bool transOCLBuiltinsFromVariables();
+  bool transOCLBuiltinFromVariable(GlobalVariable *GV,
+      SPIRVBuiltinVariableKind Kind);
+  MDString *transOCLKernelArgTypeName(SPIRVFunctionParameter *);
+
+  Value *mapFunction(SPIRVFunction *BF, Function *F) {
+    SPIRVDBG(spvdbgs() << "[mapFunction] " << *BF << " -> ";
+      dbgs() << *F << '\n';)
+    FuncMap[BF] = F;
+    return F;
+  }
+
+  Value *getTranslatedValue(SPIRVValue *BV);
+  Type *getTranslatedType(SPIRVType *BT);
+
+  SPIRVErrorLog &getErrorLog() {
+    return BM->getErrorLog();
+  }
+
+  void setCallingConv(CallInst *Call) {
+    Function *F = Call->getCalledFunction();
     assert(F);
-    Call->setCallingConv(F->getCallingConv());
-  }
-
-  void setAttrByCalledFunc(CallInst *Call);
-  Type *transFPType(SPIRVType* T);
-  BinaryOperator *transShiftLogicalBitwiseInst(SPIRVValue* BV, BasicBlock* BB,
-      Function* F);
-  void transFlags(llvm::Value* V);
-  Instruction *transCmpInst(SPIRVValue* BV, BasicBlock* BB, Function* F);
-  void transOCLBuiltinFromInstPreproc(SPIRVInstruction* BI, Type *&RetTy,
-      std::vector<SPIRVValue *> &Args);
-  Instruction* transOCLBuiltinPostproc(SPIRVInstruction* BI,
-      CallInst* CI, BasicBlock* BB, const std::string &DemangledName);
-  std::string transOCLImageTypeName(SPIRV::SPIRVTypeImage* ST);
-  std::string transOCLSampledImageTypeName(SPIRV::SPIRVTypeSampledImage* ST);
-  std::string transOCLPipeTypeName(SPIRV::SPIRVTypePipe* ST);
-  std::string transOCLImageTypeAccessQualifier(SPIRV::SPIRVTypeImage* ST);
-  std::string transOCLPipeTypeAccessQualifier(SPIRV::SPIRVTypePipe* ST);
-
-  Value *oclTransConstantSampler(SPIRV::SPIRVConstantSampler* BCS);
-  void setName(llvm::Value* V, SPIRVValue* BV);
-  template<class Source, class Func>
-  bool foreachFuncCtlMask(Source, Func);
-  llvm::GlobalValue::LinkageTypes transLinkageType(const SPIRVValue* V);
-  Instruction *transOCLAllAny(SPIRVInstruction* BI, BasicBlock *BB);
-  Instruction *transOCLRelational(SPIRVInstruction* BI, BasicBlock *BB);
-};
-
-Type *
-SPIRVToLLVM::getTranslatedType(SPIRVType *BV){
-  auto Loc = TypeMap.find(BV);
-  if (Loc != TypeMap.end())
-    return Loc->second;
-  return nullptr;
-}
-
-Value *
-SPIRVToLLVM::getTranslatedValue(SPIRVValue *BV){
-  auto Loc = ValueMap.find(BV);
-  if (Loc != ValueMap.end())
-    return Loc->second;
-  return nullptr;
-}
-
-void
-SPIRVToLLVM::setAttrByCalledFunc(CallInst *Call) {
-  Function *F = Call->getCalledFunction();
+    Call->setCallingConv(F->getCallingConv());
+  }
+
+  void setAttrByCalledFunc(CallInst *Call);
+  Type *transFPType(SPIRVType* T);
+  BinaryOperator *transShiftLogicalBitwiseInst(SPIRVValue* BV, BasicBlock* BB,
+      Function* F);
+  void transFlags(llvm::Value* V);
+  Instruction *transCmpInst(SPIRVValue* BV, BasicBlock* BB, Function* F);
+  void transOCLBuiltinFromInstPreproc(SPIRVInstruction* BI, Type *&RetTy,
+      std::vector<SPIRVValue *> &Args);
+  Instruction* transOCLBuiltinPostproc(SPIRVInstruction* BI,
+      CallInst* CI, BasicBlock* BB, const std::string &DemangledName);
+  std::string transOCLImageTypeName(SPIRV::SPIRVTypeImage* ST);
+  std::string transOCLSampledImageTypeName(SPIRV::SPIRVTypeSampledImage* ST);
+  std::string transOCLPipeTypeName(SPIRV::SPIRVTypePipe* ST);
+  std::string transOCLImageTypeAccessQualifier(SPIRV::SPIRVTypeImage* ST);
+  std::string transOCLPipeTypeAccessQualifier(SPIRV::SPIRVTypePipe* ST);
+
+  Value *oclTransConstantSampler(SPIRV::SPIRVConstantSampler* BCS);
+  void setName(llvm::Value* V, SPIRVValue* BV);
+  template<class Source, class Func>
+  bool foreachFuncCtlMask(Source, Func);
+  llvm::GlobalValue::LinkageTypes transLinkageType(const SPIRVValue* V);
+  Instruction *transOCLAllAny(SPIRVInstruction* BI, BasicBlock *BB);
+  Instruction *transOCLRelational(SPIRVInstruction* BI, BasicBlock *BB);
+};
+
+Type *
+SPIRVToLLVM::getTranslatedType(SPIRVType *BV){
+  auto Loc = TypeMap.find(BV);
+  if (Loc != TypeMap.end())
+    return Loc->second;
+  return nullptr;
+}
+
+Value *
+SPIRVToLLVM::getTranslatedValue(SPIRVValue *BV){
+  auto Loc = ValueMap.find(BV);
+  if (Loc != ValueMap.end())
+    return Loc->second;
+  return nullptr;
+}
+
+void
+SPIRVToLLVM::setAttrByCalledFunc(CallInst *Call) {
+  Function *F = Call->getCalledFunction();
   assert(F);
-  if (F->isIntrinsic()) {
-    return;
-  }
-  Call->setCallingConv(F->getCallingConv());
-  Call->setAttributes(F->getAttributes());
-}
-
-bool
-SPIRVToLLVM::transOCLBuiltinsFromVariables(){
-  std::vector<GlobalVariable *> WorkList;
-  for (auto I = M->global_begin(), E = M->global_end(); I != E; ++I) {
-    SPIRVBuiltinVariableKind Kind;
-    if (!isSPIRVBuiltinVariable(I, &Kind))
-      continue;
-    if (!transOCLBuiltinFromVariable(I, Kind))
-      return false;
-    WorkList.push_back(I);
-  }
-  for (auto &I:WorkList) {
-    I->dropAllReferences();
-    I->removeFromParent();
-  }
-  return true;
-}
-
-// For integer types shorter than 32 bit, unsigned/signedness can be inferred
-// from zext/sext attribute.
-MDString *
-SPIRVToLLVM::transOCLKernelArgTypeName(SPIRVFunctionParameter *Arg) {
-  auto Ty = Arg->isByVal() ? Arg->getType()->getPointerElementType() :
-    Arg->getType();
-  return MDString::get(*Context, transTypeToOCLTypeName(Ty, !Arg->isZext()));
-}
-
-// Variable like GlobalInvolcationId[x] -> get_global_id(x).
-// Variable like WorkDim -> get_work_dim().
-bool
-SPIRVToLLVM::transOCLBuiltinFromVariable(GlobalVariable *GV,
-    SPIRVBuiltinVariableKind Kind) {
-  std::string FuncName = SPIRSPIRVBuiltinVariableMap::rmap(Kind);
-  std::string MangledName;
-  Type *ReturnTy =  GV->getType()->getPointerElementType();
-  bool IsVec = ReturnTy->isVectorTy();
-  if (IsVec)
-    ReturnTy = cast<VectorType>(ReturnTy)->getElementType();
-  std::vector<Type*> ArgTy;
-  if (IsVec)
-    ArgTy.push_back(Type::getInt32Ty(*Context));
-  MangleOpenCLBuiltin(FuncName, ArgTy, MangledName);
-  Function *Func = M->getFunction(MangledName);
-  if (!Func) {
-    FunctionType *FT = FunctionType::get(ReturnTy, ArgTy, false);
-    Func = Function::Create(FT, GlobalValue::ExternalLinkage, MangledName, M);
-    Func->setCallingConv(CallingConv::SPIR_FUNC);
-    Func->addFnAttr(Attribute::NoUnwind);
-    Func->addFnAttr(Attribute::ReadNone);
-  }
-  std::vector<Instruction *> Deletes;
-  std::vector<Instruction *> Uses;
-  for (auto UI = GV->user_begin(), UE = GV->user_end(); UI != UE; ++UI) {
-    assert (isa<LoadInst>(*UI) && "Unsupported use");
-    auto LD = dyn_cast<LoadInst>(*UI);
-    if (!IsVec) {
-      Uses.push_back(LD);
-      Deletes.push_back(LD);
-      continue;
-    }
-    for (auto LDUI = LD->user_begin(), LDUE = LD->user_end(); LDUI != LDUE;
-        ++LDUI) {
-      assert(isa<ExtractElementInst>(*LDUI) && "Unsupported use");
-      auto EEI = dyn_cast<ExtractElementInst>(*LDUI);
-      Uses.push_back(EEI);
-      Deletes.push_back(EEI);
-    }
-    Deletes.push_back(LD);
-  }
-  for (auto &I:Uses) {
-    std::vector<Value *> Arg;
-    if (auto EEI = dyn_cast<ExtractElementInst>(I))
-      Arg.push_back(EEI->getIndexOperand());
-    auto Call = CallInst::Create(Func, Arg, "", I);
-    Call->takeName(I);
-    setAttrByCalledFunc(Call);
-    SPIRVDBG(dbgs() << "[transOCLBuiltinFromVariable] " << *I << " -> " <<
-        *Call << '\n';)
-    I->replaceAllUsesWith(Call);
-  }
-  for (auto &I:Deletes) {
-    I->dropAllReferences();
-    I->removeFromParent();
-  }
-  return true;
-}
-
-Type *
-SPIRVToLLVM::transFPType(SPIRVType* T) {
-  switch(T->getFloatBitWidth()) {
-  case 16: return Type::getHalfTy(*Context);
-  case 32: return Type::getFloatTy(*Context);
-  case 64: return Type::getDoubleTy(*Context);
-  default:
-    llvm_unreachable("Invalid type");
-    return nullptr;
-  }
-}
-
-std::string
-SPIRVToLLVM::transOCLImageTypeName(SPIRV::SPIRVTypeImage* ST) {
-  std::string Name = std::string(kSPR2TypeName::OCLPrefix)
-    + rmap<std::string>(ST->getDescriptor());
-  if (SPIRVGenImgTypeAccQualPostfix)
-    Name = Name + kSPR2TypeName::Delimiter
-      + rmap<std::string>(ST->getAccessQualifier());
-  return std::move(Name);
-}
-
-std::string
-SPIRVToLLVM::transOCLSampledImageTypeName(SPIRV::SPIRVTypeSampledImage* ST) {
-  return getSPIRVTypeName(kSPIRVTypeName::SampledImg,
-    getSPIRVImageTypePostfixes(getSPIRVImageSampledTypeName(
-      ST->getImageType()->getSampledType()),
-      ST->getImageType()->getDescriptor(),
-      ST->getImageType()->getAccessQualifier()));
-}
-
-std::string
-SPIRVToLLVM::transOCLPipeTypeName(SPIRV::SPIRVTypePipe* PT) {
-  return kSPR2TypeName::Pipe;
-}
-
-Type *
-SPIRVToLLVM::transType(SPIRVType *T) {
-  auto Loc = TypeMap.find(T);
-  if (Loc != TypeMap.end())
-    return Loc->second;
-
-  SPIRVDBG(spvdbgs() << "[transType] " << *T << " -> ";)
-  T->validate();
-  switch(T->getOpCode()) {
-  case OpTypeVoid:
-    return mapType(T, Type::getVoidTy(*Context));
-  case OpTypeBool:
-    return mapType(T, Type::getInt1Ty(*Context));
-  case OpTypeInt:
-    return mapType(T, Type::getIntNTy(*Context, T->getIntegerBitWidth()));
-  case OpTypeFloat:
-    return mapType(T, transFPType(T));
-  case OpTypeArray:
-    return mapType(T, ArrayType::get(transType(T->getArrayElementType()),
-        T->getArrayLength()));
-  case OpTypePointer:
-    return mapType(T, PointerType::get(transType(T->getPointerElementType()),
-        SPIRSPIRVAddrSpaceMap::rmap(T->getPointerStorageClass())));
-  case OpTypeVector:
-    return mapType(T, VectorType::get(transType(T->getVectorComponentType()),
-        T->getVectorComponentCount()));
-  case OpTypeOpaque:
-    return mapType(T, StructType::create(*Context, T->getName()));
-  case OpTypeFunction: {
-    auto FT = static_cast<SPIRVTypeFunction *>(T);
-    auto RT = transType(FT->getReturnType());
-    std::vector<Type *> PT;
-    for (size_t I = 0, E = FT->getNumParameters(); I != E; ++I)
-      PT.push_back(transType(FT->getParameterType(I)));
-    return mapType(T, FunctionType::get(RT, PT, false));
-    }
-  case OpTypeImage: {
-    auto ST = static_cast<SPIRVTypeImage *>(T);
-    if (ST->isOCLImage())
-      return mapType(T, getOrCreateOpaquePtrType(M,
-          transOCLImageTypeName(ST)));
-    else
-      llvm_unreachable("Unsupported image type");
-    return nullptr;
-  }
-  case OpTypeSampler:
-    return mapType(T, Type::getInt32Ty(*Context));
-  case OpTypeSampledImage: {
-    auto ST = static_cast<SPIRVTypeSampledImage *>(T);
-    return mapType(T, getOrCreateOpaquePtrType(M,
-        transOCLSampledImageTypeName(ST)));
-  }
-  case OpTypeStruct: {
-    auto ST = static_cast<SPIRVTypeStruct *>(T);
-    auto Name = ST->getName();
-    if (!Name.empty()) {
-      if (auto OldST = M->getTypeByName(Name))
-        OldST->setName("");
-    }
-    auto *StructTy = StructType::create(*Context, Name);
-    mapType(ST, StructTy);
-    SmallVector<Type *, 4> MT;
-    for (size_t I = 0, E = ST->getMemberCount(); I != E; ++I)
-      MT.push_back(transType(ST->getMemberType(I)));
-    StructTy->setBody(MT, ST->isPacked());
-    return StructTy;
-  }
-  case OpTypePipe: {
-    auto PT = static_cast<SPIRVTypePipe *>(T);
-    return mapType(T, getOrCreateOpaquePtrType(M,
-        transOCLPipeTypeName(PT),
-        getOCLOpaqueTypeAddrSpace(T->getOpCode())));
-    }
-  default: {
-    auto OC = T->getOpCode();
-    if (isOpaqueGenericTypeOpCode(OC))
-      return mapType(T, getOrCreateOpaquePtrType(M,
-          OCLOpaqueTypeOpCodeMap::rmap(OC),
-          getOCLOpaqueTypeAddrSpace(OC)));
-    llvm_unreachable("Not implemented");
-    }
-  }
-  return 0;
-}
-
-std::string
-SPIRVToLLVM::transTypeToOCLTypeName(SPIRVType *T, bool IsSigned) {
-  switch(T->getOpCode()) {
-  case OpTypeVoid:
-    return "void";
-  case OpTypeBool:
-    return "bool";
-  case OpTypeInt: {
-    std::string Prefix = IsSigned ? "" : "u";
-    switch(T->getIntegerBitWidth()) {
-    case 8:
-      return Prefix + "char";
-    case 16:
-      return Prefix + "short";
-    case 32:
-      return Prefix + "int";
-    case 64:
-      return Prefix + "long";
-    default:
-      llvm_unreachable("invalid integer size");
-      return Prefix + std::string("int") + T->getIntegerBitWidth() + "_t";
-    }
-  }
-  break;
-  case OpTypeFloat:
-    switch(T->getFloatBitWidth()){
-    case 16:
-      return "half";
-    case 32:
-      return "float";
-    case 64:
-      return "double";
-    default:
-      llvm_unreachable("invalid floating pointer bitwidth");
-      return std::string("float") + T->getFloatBitWidth() + "_t";
-    }
-    break;
-  case OpTypeArray:
-    return "array";
-  case OpTypePointer:
-    return transTypeToOCLTypeName(T->getPointerElementType()) + "*";
-  case OpTypeVector:
-    return transTypeToOCLTypeName(T->getVectorComponentType()) +
-        T->getVectorComponentCount();
-  case OpTypeOpaque:
-      return T->getName();
-  case OpTypeFunction:
-    llvm_unreachable("Unsupported");
-    return "function";
-  case OpTypeStruct: {
-    auto Name = T->getName();
-    if (Name.find("struct.") == 0)
-      Name[6] = ' ';
-    else if (Name.find("union.") == 0)
-      Name[5] = ' ';
-    return Name;
-  }
-  case OpTypePipe:
-    return "pipe";
-  case OpTypeSampler:
-    return "sampler_t";
-  case OpTypeImage:
-    return rmap<std::string>(static_cast<SPIRVTypeImage *>(T)->getDescriptor());
-  default:
-      if (isOpaqueGenericTypeOpCode(T->getOpCode())) {
-        return OCLOpaqueTypeOpCodeMap::rmap(T->getOpCode());
-      }
-      llvm_unreachable("Not implemented");
-      return "unknown";
-  }
-}
-
-std::vector<Type *>
-SPIRVToLLVM::transTypeVector(const std::vector<SPIRVType *> &BT) {
-  std::vector<Type *> T;
-  for (auto I: BT)
-    T.push_back(transType(I));
-  return T;
-}
-
-std::vector<Value *>
-SPIRVToLLVM::transValue(const std::vector<SPIRVValue *> &BV, Function *F,
-    BasicBlock *BB) {
-  std::vector<Value *> V;
-  for (auto I: BV)
-    V.push_back(transValue(I, F, BB));
-  return V;
-}
-
-bool
-SPIRVToLLVM::isSPIRVCmpInstTransToLLVMInst(SPIRVInstruction* BI) const {
-  auto OC = BI->getOpCode();
-  return isCmpOpCode(OC) &&
-      !(OC >= OpLessOrGreater && OC <= OpUnordered);
-}
-
-void
-SPIRVToLLVM::transFlags(llvm::Value* V) {
-  if(!isa<Instruction>(V))
-    return;
-  auto OC = cast<Instruction>(V)->getOpcode();
-  if (OC == Instruction::AShr || OC == Instruction::LShr) {
-    cast<BinaryOperator>(V)->setIsExact();
-    return;
-  }
-}
-
-void
-SPIRVToLLVM::setName(llvm::Value* V, SPIRVValue* BV) {
-  auto Name = BV->getName();
-  if (!Name.empty() && (!V->hasName() || Name != V->getName()))
-    V->setName(Name);
-}
-
-Value *
-SPIRVToLLVM::transValue(SPIRVValue *BV, Function *F, BasicBlock *BB,
-    bool CreatePlaceHolder){
-  SPIRVToLLVMValueMap::iterator Loc = ValueMap.find(BV);
-  if (Loc != ValueMap.end() && (!PlaceholderMap.count(BV) || CreatePlaceHolder))
-    return Loc->second;
-
-  SPIRVDBG(spvdbgs() << "[transValue] " << *BV << " -> ";)
-  BV->validate();
-
-  auto V = transValueWithoutDecoration(BV, F, BB, CreatePlaceHolder);
-  if (!V) {
-    SPIRVDBG(dbgs() << " Warning ! nullptr\n";)
-    return nullptr;
-  }
-  setName(V, BV);
-  if (!transDecoration(BV, V)) {
-    assert (0 && "trans decoration fail");
-    return nullptr;
-  }
-  transFlags(V);
-
-  SPIRVDBG(dbgs() << *V << '\n';)
-
-  return V;
-}
-
-Value *
-SPIRVToLLVM::transConvertInst(SPIRVValue* BV, Function* F, BasicBlock* BB) {
-  SPIRVUnary* BC = static_cast<SPIRVUnary*>(BV);
-  auto Src = transValue(BC->getOperand(0), F, BB, BB ? true : false);
-  auto Dst = transType(BC->getType());
-  CastInst::CastOps CO = Instruction::BitCast;
-  bool IsExt = Dst->getScalarSizeInBits()
-      > Src->getType()->getScalarSizeInBits();
-  switch (BC->getOpCode()) {
-  case OpPtrCastToGeneric:
-  case OpGenericCastToPtr:
-    CO = Instruction::AddrSpaceCast;
-    break;
-  case OpSConvert:
-    CO = IsExt ? Instruction::SExt : Instruction::Trunc;
-    break;
-  case OpUConvert:
-    CO = IsExt ? Instruction::ZExt : Instruction::Trunc;
-    break;
-  case OpFConvert:
-    CO = IsExt ? Instruction::FPExt : Instruction::FPTrunc;
-    break;
-  default:
-    CO = static_cast<CastInst::CastOps>(OpCodeMap::rmap(BC->getOpCode()));
-  }
-  assert(CastInst::isCast(CO) && "Invalid cast op code");
-  SPIRVDBG(if (!CastInst::castIsValid(CO, Src, Dst)) {
-    spvdbgs() << "Invalid cast: " << *BV << " -> ";
-    dbgs() << "Op = " << CO << ", Src = " << *Src << " Dst = " << *Dst << '\n';
-  })
-  if (BB)
-    return CastInst::Create(CO, Src, Dst, BV->getName(), BB);
-  return ConstantExpr::getCast(CO, dyn_cast<Constant>(Src), Dst);
-}
-
-BinaryOperator *SPIRVToLLVM::transShiftLogicalBitwiseInst(SPIRVValue* BV,
-    BasicBlock* BB,Function* F) {
-  SPIRVBinary* BBN = static_cast<SPIRVBinary*>(BV);
-  assert(BB && "Invalid BB");
-  Instruction::BinaryOps BO;
-  auto OP = BBN->getOpCode();
-  if (isLogicalOpCode(OP))
-    OP = IntBoolOpMap::rmap(OP);
-  BO = static_cast<Instruction::BinaryOps>(OpCodeMap::rmap(OP));
-  auto Inst = BinaryOperator::Create(BO,
-      transValue(BBN->getOperand(0), F, BB),
-      transValue(BBN->getOperand(1), F, BB), BV->getName(), BB);
-  return Inst;
-}
-
-Instruction *
-SPIRVToLLVM::transCmpInst(SPIRVValue* BV, BasicBlock* BB, Function* F) {
-  SPIRVCompare* BC = static_cast<SPIRVCompare*>(BV);
-  assert(BB && "Invalid BB");
-  SPIRVType* BT = BC->getOperand(0)->getType();
-  Instruction* Inst = nullptr;
-  auto OP = BC->getOpCode();
-  if (isLogicalOpCode(OP))
-    OP = IntBoolOpMap::rmap(OP);
-  if (BT->isTypeVectorOrScalarInt() || BT->isTypeVectorOrScalarBool() ||
-      BT->isTypePointer())
-    Inst = new ICmpInst(*BB, CmpMap::rmap(OP),
-        transValue(BC->getOperand(0), F, BB),
-        transValue(BC->getOperand(1), F, BB));
-  else if (BT->isTypeVectorOrScalarFloat())
-    Inst = new FCmpInst(*BB, CmpMap::rmap(OP),
-        transValue(BC->getOperand(0), F, BB),
-        transValue(BC->getOperand(1), F, BB));
-  assert(Inst && "not implemented");
-  return Inst;
-}
-
-bool
-SPIRVToLLVM::postProcessOCL() {
-  std::string DemangledName;
-  SPIRVWord SrcLangVer = 0;
-  BM->getSourceLanguage(&SrcLangVer);
-  bool isCPP = SrcLangVer == kOCLVer::CL21;
-  for (auto I = M->begin(), E = M->end(); I != E;) {
-    auto F = I++;
-    if (F->hasName() && F->isDeclaration()) {
-      DEBUG(dbgs() << "[postProcessOCL sret] " << *F << '\n');
-      if (F->getReturnType()->isStructTy() &&
-          oclIsBuiltin(F->getName(), &DemangledName, isCPP)) {
-        if (!postProcessOCLBuiltinReturnStruct(F))
-          return false;
-      }
-    }
-  }
-  for (auto I = M->begin(), E = M->end(); I != E;) {
-    auto F = I++;
-    if (F->hasName() && F->isDeclaration()) {
-      DEBUG(dbgs() << "[postProcessOCL func ptr] " << *F << '\n');
-      auto AI = F->arg_begin();
-      if (hasFunctionPointerArg(F, AI) && isDecoratedSPIRVFunc(F))
-        if (!postProcessOCLBuiltinWithFuncPointer(F, AI))
-          return false;
-    }
-  }
-  for (auto I = M->begin(), E = M->end(); I != E;) {
-    auto F = I++;
-    if (F->hasName() && F->isDeclaration()) {
-      DEBUG(dbgs() << "[postProcessOCL array arg] " << *F << '\n');
-      if (hasArrayArg(F) && oclIsBuiltin(F->getName(), &DemangledName, isCPP))
-        if (!postProcessOCLBuiltinWithArrayArguments(F, DemangledName))
-          return false;
-    }
-  }
-  return true;
-}
-
-bool
-SPIRVToLLVM::postProcessOCLBuiltinReturnStruct(Function *F) {
-  std::string Name = F->getName();
-  F->setName(Name + ".old");
-  for (auto I = F->user_begin(), E = F->user_end(); I != E;) {
-    if (auto CI = dyn_cast<CallInst>(*I++)) {
-      auto ST = dyn_cast<StoreInst>(*(CI->user_begin()));
+  if (F->isIntrinsic()) {
+    return;
+  }
+  Call->setCallingConv(F->getCallingConv());
+  Call->setAttributes(F->getAttributes());
+}
+
+bool
+SPIRVToLLVM::transOCLBuiltinsFromVariables(){
+  std::vector<GlobalVariable *> WorkList;
+  for (auto I = M->global_begin(), E = M->global_end(); I != E; ++I) {
+    SPIRVBuiltinVariableKind Kind;
+    if (!isSPIRVBuiltinVariable(&*I, &Kind))
+      continue;
+    if (!transOCLBuiltinFromVariable(&*I, Kind))
+      return false;
+    WorkList.push_back(&*I);
+  }
+  for (auto &I:WorkList) {
+    I->dropAllReferences();
+    I->removeFromParent();
+  }
+  return true;
+}
+
+// For integer types shorter than 32 bit, unsigned/signedness can be inferred
+// from zext/sext attribute.
+MDString *
+SPIRVToLLVM::transOCLKernelArgTypeName(SPIRVFunctionParameter *Arg) {
+  auto Ty = Arg->isByVal() ? Arg->getType()->getPointerElementType() :
+    Arg->getType();
+  return MDString::get(*Context, transTypeToOCLTypeName(Ty, !Arg->isZext()));
+}
+
+// Variable like GlobalInvolcationId[x] -> get_global_id(x).
+// Variable like WorkDim -> get_work_dim().
+bool
+SPIRVToLLVM::transOCLBuiltinFromVariable(GlobalVariable *GV,
+    SPIRVBuiltinVariableKind Kind) {
+  std::string FuncName = SPIRSPIRVBuiltinVariableMap::rmap(Kind);
+  std::string MangledName;
+  Type *ReturnTy =  GV->getType()->getPointerElementType();
+  bool IsVec = ReturnTy->isVectorTy();
+  if (IsVec)
+    ReturnTy = cast<VectorType>(ReturnTy)->getElementType();
+  std::vector<Type*> ArgTy;
+  if (IsVec)
+    ArgTy.push_back(Type::getInt32Ty(*Context));
+  MangleOpenCLBuiltin(FuncName, ArgTy, MangledName);
+  Function *Func = M->getFunction(MangledName);
+  if (!Func) {
+    FunctionType *FT = FunctionType::get(ReturnTy, ArgTy, false);
+    Func = Function::Create(FT, GlobalValue::ExternalLinkage, MangledName, M);
+    Func->setCallingConv(CallingConv::SPIR_FUNC);
+    Func->addFnAttr(Attribute::NoUnwind);
+    Func->addFnAttr(Attribute::ReadNone);
+  }
+  std::vector<Instruction *> Deletes;
+  std::vector<Instruction *> Uses;
+  for (auto UI = GV->user_begin(), UE = GV->user_end(); UI != UE; ++UI) {
+    assert (isa<LoadInst>(*UI) && "Unsupported use");
+    auto LD = dyn_cast<LoadInst>(*UI);
+    if (!IsVec) {
+      Uses.push_back(LD);
+      Deletes.push_back(LD);
+      continue;
+    }
+    for (auto LDUI = LD->user_begin(), LDUE = LD->user_end(); LDUI != LDUE;
+        ++LDUI) {
+      assert(isa<ExtractElementInst>(*LDUI) && "Unsupported use");
+      auto EEI = dyn_cast<ExtractElementInst>(*LDUI);
+      Uses.push_back(EEI);
+      Deletes.push_back(EEI);
+    }
+    Deletes.push_back(LD);
+  }
+  for (auto &I:Uses) {
+    std::vector<Value *> Arg;
+    if (auto EEI = dyn_cast<ExtractElementInst>(I))
+      Arg.push_back(EEI->getIndexOperand());
+    auto Call = CallInst::Create(Func, Arg, "", I);
+    Call->takeName(I);
+    setAttrByCalledFunc(Call);
+    SPIRVDBG(dbgs() << "[transOCLBuiltinFromVariable] " << *I << " -> " <<
+        *Call << '\n';)
+    I->replaceAllUsesWith(Call);
+  }
+  for (auto &I:Deletes) {
+    I->dropAllReferences();
+    I->removeFromParent();
+  }
+  return true;
+}
+
+Type *
+SPIRVToLLVM::transFPType(SPIRVType* T) {
+  switch(T->getFloatBitWidth()) {
+  case 16: return Type::getHalfTy(*Context);
+  case 32: return Type::getFloatTy(*Context);
+  case 64: return Type::getDoubleTy(*Context);
+  default:
+    llvm_unreachable("Invalid type");
+    return nullptr;
+  }
+}
+
+std::string
+SPIRVToLLVM::transOCLImageTypeName(SPIRV::SPIRVTypeImage* ST) {
+  std::string Name = std::string(kSPR2TypeName::OCLPrefix)
+    + rmap<std::string>(ST->getDescriptor());
+  if (SPIRVGenImgTypeAccQualPostfix)
+    Name = Name + kSPR2TypeName::Delimiter
+      + rmap<std::string>(ST->getAccessQualifier());
+  return std::move(Name);
+}
+
+std::string
+SPIRVToLLVM::transOCLSampledImageTypeName(SPIRV::SPIRVTypeSampledImage* ST) {
+  return getSPIRVTypeName(kSPIRVTypeName::SampledImg,
+    getSPIRVImageTypePostfixes(getSPIRVImageSampledTypeName(
+      ST->getImageType()->getSampledType()),
+      ST->getImageType()->getDescriptor(),
+      ST->getImageType()->getAccessQualifier()));
+}
+
+std::string
+SPIRVToLLVM::transOCLPipeTypeName(SPIRV::SPIRVTypePipe* PT) {
+  return kSPR2TypeName::Pipe;
+}
+
+Type *
+SPIRVToLLVM::transType(SPIRVType *T) {
+  auto Loc = TypeMap.find(T);
+  if (Loc != TypeMap.end())
+    return Loc->second;
+
+  SPIRVDBG(spvdbgs() << "[transType] " << *T << " -> ";)
+  T->validate();
+  switch(T->getOpCode()) {
+  case OpTypeVoid:
+    return mapType(T, Type::getVoidTy(*Context));
+  case OpTypeBool:
+    return mapType(T, Type::getInt1Ty(*Context));
+  case OpTypeInt:
+    return mapType(T, Type::getIntNTy(*Context, T->getIntegerBitWidth()));
+  case OpTypeFloat:
+    return mapType(T, transFPType(T));
+  case OpTypeArray:
+    return mapType(T, ArrayType::get(transType(T->getArrayElementType()),
+        T->getArrayLength()));
+  case OpTypePointer:
+    return mapType(T, PointerType::get(transType(T->getPointerElementType()),
+        SPIRSPIRVAddrSpaceMap::rmap(T->getPointerStorageClass())));
+  case OpTypeVector:
+    return mapType(T, VectorType::get(transType(T->getVectorComponentType()),
+        T->getVectorComponentCount()));
+  case OpTypeOpaque:
+    return mapType(T, StructType::create(*Context, T->getName()));
+  case OpTypeFunction: {
+    auto FT = static_cast<SPIRVTypeFunction *>(T);
+    auto RT = transType(FT->getReturnType());
+    std::vector<Type *> PT;
+    for (size_t I = 0, E = FT->getNumParameters(); I != E; ++I)
+      PT.push_back(transType(FT->getParameterType(I)));
+    return mapType(T, FunctionType::get(RT, PT, false));
+    }
+  case OpTypeImage: {
+    auto ST = static_cast<SPIRVTypeImage *>(T);
+    if (ST->isOCLImage())
+      return mapType(T, getOrCreateOpaquePtrType(M,
+          transOCLImageTypeName(ST)));
+    else
+      llvm_unreachable("Unsupported image type");
+    return nullptr;
+  }
+  case OpTypeSampler:
+    return mapType(T, Type::getInt32Ty(*Context));
+  case OpTypeSampledImage: {
+    auto ST = static_cast<SPIRVTypeSampledImage *>(T);
+    return mapType(T, getOrCreateOpaquePtrType(M,
+        transOCLSampledImageTypeName(ST)));
+  }
+  case OpTypeStruct: {
+    auto ST = static_cast<SPIRVTypeStruct *>(T);
+    auto Name = ST->getName();
+    if (!Name.empty()) {
+      if (auto OldST = M->getTypeByName(Name))
+        OldST->setName("");
+    }
+    auto *StructTy = StructType::create(*Context, Name);
+    mapType(ST, StructTy);
+    SmallVector<Type *, 4> MT;
+    for (size_t I = 0, E = ST->getMemberCount(); I != E; ++I)
+      MT.push_back(transType(ST->getMemberType(I)));
+    StructTy->setBody(MT, ST->isPacked());
+    return StructTy;
+  }
+  case OpTypePipe: {
+    auto PT = static_cast<SPIRVTypePipe *>(T);
+    return mapType(T, getOrCreateOpaquePtrType(M,
+        transOCLPipeTypeName(PT),
+        getOCLOpaqueTypeAddrSpace(T->getOpCode())));
+    }
+  default: {
+    auto OC = T->getOpCode();
+    if (isOpaqueGenericTypeOpCode(OC))
+      return mapType(T, getOrCreateOpaquePtrType(M,
+          OCLOpaqueTypeOpCodeMap::rmap(OC),
+          getOCLOpaqueTypeAddrSpace(OC)));
+    llvm_unreachable("Not implemented");
+    }
+  }
+  return 0;
+}
+
+std::string
+SPIRVToLLVM::transTypeToOCLTypeName(SPIRVType *T, bool IsSigned) {
+  switch(T->getOpCode()) {
+  case OpTypeVoid:
+    return "void";
+  case OpTypeBool:
+    return "bool";
+  case OpTypeInt: {
+    std::string Prefix = IsSigned ? "" : "u";
+    switch(T->getIntegerBitWidth()) {
+    case 8:
+      return Prefix + "char";
+    case 16:
+      return Prefix + "short";
+    case 32:
+      return Prefix + "int";
+    case 64:
+      return Prefix + "long";
+    default:
+      llvm_unreachable("invalid integer size");
+      return Prefix + std::string("int") + T->getIntegerBitWidth() + "_t";
+    }
+  }
+  break;
+  case OpTypeFloat:
+    switch(T->getFloatBitWidth()){
+    case 16:
+      return "half";
+    case 32:
+      return "float";
+    case 64:
+      return "double";
+    default:
+      llvm_unreachable("invalid floating pointer bitwidth");
+      return std::string("float") + T->getFloatBitWidth() + "_t";
+    }
+    break;
+  case OpTypeArray:
+    return "array";
+  case OpTypePointer:
+    return transTypeToOCLTypeName(T->getPointerElementType()) + "*";
+  case OpTypeVector:
+    return transTypeToOCLTypeName(T->getVectorComponentType()) +
+        T->getVectorComponentCount();
+  case OpTypeOpaque:
+      return T->getName();
+  case OpTypeFunction:
+    llvm_unreachable("Unsupported");
+    return "function";
+  case OpTypeStruct: {
+    auto Name = T->getName();
+    if (Name.find("struct.") == 0)
+      Name[6] = ' ';
+    else if (Name.find("union.") == 0)
+      Name[5] = ' ';
+    return Name;
+  }
+  case OpTypePipe:
+    return "pipe";
+  case OpTypeSampler:
+    return "sampler_t";
+  case OpTypeImage:
+    return rmap<std::string>(static_cast<SPIRVTypeImage *>(T)->getDescriptor());
+  default:
+      if (isOpaqueGenericTypeOpCode(T->getOpCode())) {
+        return OCLOpaqueTypeOpCodeMap::rmap(T->getOpCode());
+      }
+      llvm_unreachable("Not implemented");
+      return "unknown";
+  }
+}
+
+std::vector<Type *>
+SPIRVToLLVM::transTypeVector(const std::vector<SPIRVType *> &BT) {
+  std::vector<Type *> T;
+  for (auto I: BT)
+    T.push_back(transType(I));
+  return T;
+}
+
+std::vector<Value *>
+SPIRVToLLVM::transValue(const std::vector<SPIRVValue *> &BV, Function *F,
+    BasicBlock *BB) {
+  std::vector<Value *> V;
+  for (auto I: BV)
+    V.push_back(transValue(I, F, BB));
+  return V;
+}
+
+bool
+SPIRVToLLVM::isSPIRVCmpInstTransToLLVMInst(SPIRVInstruction* BI) const {
+  auto OC = BI->getOpCode();
+  return isCmpOpCode(OC) &&
+      !(OC >= OpLessOrGreater && OC <= OpUnordered);
+}
+
+void
+SPIRVToLLVM::transFlags(llvm::Value* V) {
+  if(!isa<Instruction>(V))
+    return;
+  auto OC = cast<Instruction>(V)->getOpcode();
+  if (OC == Instruction::AShr || OC == Instruction::LShr) {
+    cast<BinaryOperator>(V)->setIsExact();
+    return;
+  }
+}
+
+void
+SPIRVToLLVM::setName(llvm::Value* V, SPIRVValue* BV) {
+  auto Name = BV->getName();
+  if (!Name.empty() && (!V->hasName() || Name != V->getName()))
+    V->setName(Name);
+}
+
+Value *
+SPIRVToLLVM::transValue(SPIRVValue *BV, Function *F, BasicBlock *BB,
+    bool CreatePlaceHolder){
+  SPIRVToLLVMValueMap::iterator Loc = ValueMap.find(BV);
+  if (Loc != ValueMap.end() && (!PlaceholderMap.count(BV) || CreatePlaceHolder))
+    return Loc->second;
+
+  SPIRVDBG(spvdbgs() << "[transValue] " << *BV << " -> ";)
+  BV->validate();
+
+  auto V = transValueWithoutDecoration(BV, F, BB, CreatePlaceHolder);
+  if (!V) {
+    SPIRVDBG(dbgs() << " Warning ! nullptr\n";)
+    return nullptr;
+  }
+  setName(V, BV);
+  if (!transDecoration(BV, V)) {
+    assert (0 && "trans decoration fail");
+    return nullptr;
+  }
+  transFlags(V);
+
+  SPIRVDBG(dbgs() << *V << '\n';)
+
+  return V;
+}
+
+Value *
+SPIRVToLLVM::transConvertInst(SPIRVValue* BV, Function* F, BasicBlock* BB) {
+  SPIRVUnary* BC = static_cast<SPIRVUnary*>(BV);
+  auto Src = transValue(BC->getOperand(0), F, BB, BB ? true : false);
+  auto Dst = transType(BC->getType());
+  CastInst::CastOps CO = Instruction::BitCast;
+  bool IsExt = Dst->getScalarSizeInBits()
+      > Src->getType()->getScalarSizeInBits();
+  switch (BC->getOpCode()) {
+  case OpPtrCastToGeneric:
+  case OpGenericCastToPtr:
+    CO = Instruction::AddrSpaceCast;
+    break;
+  case OpSConvert:
+    CO = IsExt ? Instruction::SExt : Instruction::Trunc;
+    break;
+  case OpUConvert:
+    CO = IsExt ? Instruction::ZExt : Instruction::Trunc;
+    break;
+  case OpFConvert:
+    CO = IsExt ? Instruction::FPExt : Instruction::FPTrunc;
+    break;
+  default:
+    CO = static_cast<CastInst::CastOps>(OpCodeMap::rmap(BC->getOpCode()));
+  }
+  assert(CastInst::isCast(CO) && "Invalid cast op code");
+  SPIRVDBG(if (!CastInst::castIsValid(CO, Src, Dst)) {
+    spvdbgs() << "Invalid cast: " << *BV << " -> ";
+    dbgs() << "Op = " << CO << ", Src = " << *Src << " Dst = " << *Dst << '\n';
+  })
+  if (BB)
+    return CastInst::Create(CO, Src, Dst, BV->getName(), BB);
+  return ConstantExpr::getCast(CO, dyn_cast<Constant>(Src), Dst);
+}
+
+BinaryOperator *SPIRVToLLVM::transShiftLogicalBitwiseInst(SPIRVValue* BV,
+    BasicBlock* BB,Function* F) {
+  SPIRVBinary* BBN = static_cast<SPIRVBinary*>(BV);
+  assert(BB && "Invalid BB");
+  Instruction::BinaryOps BO;
+  auto OP = BBN->getOpCode();
+  if (isLogicalOpCode(OP))
+    OP = IntBoolOpMap::rmap(OP);
+  BO = static_cast<Instruction::BinaryOps>(OpCodeMap::rmap(OP));
+  auto Inst = BinaryOperator::Create(BO,
+      transValue(BBN->getOperand(0), F, BB),
+      transValue(BBN->getOperand(1), F, BB), BV->getName(), BB);
+  return Inst;
+}
+
+Instruction *
+SPIRVToLLVM::transCmpInst(SPIRVValue* BV, BasicBlock* BB, Function* F) {
+  SPIRVCompare* BC = static_cast<SPIRVCompare*>(BV);
+  assert(BB && "Invalid BB");
+  SPIRVType* BT = BC->getOperand(0)->getType();
+  Instruction* Inst = nullptr;
+  auto OP = BC->getOpCode();
+  if (isLogicalOpCode(OP))
+    OP = IntBoolOpMap::rmap(OP);
+  if (BT->isTypeVectorOrScalarInt() || BT->isTypeVectorOrScalarBool() ||
+      BT->isTypePointer())
+    Inst = new ICmpInst(*BB, CmpMap::rmap(OP),
+        transValue(BC->getOperand(0), F, BB),
+        transValue(BC->getOperand(1), F, BB));
+  else if (BT->isTypeVectorOrScalarFloat())
+    Inst = new FCmpInst(*BB, CmpMap::rmap(OP),
+        transValue(BC->getOperand(0), F, BB),
+        transValue(BC->getOperand(1), F, BB));
+  assert(Inst && "not implemented");
+  return Inst;
+}
+
+bool
+SPIRVToLLVM::postProcessOCL() {
+  std::string DemangledName;
+  SPIRVWord SrcLangVer = 0;
+  BM->getSourceLanguage(&SrcLangVer);
+  bool isCPP = SrcLangVer == kOCLVer::CL21;
+  for (auto I = M->begin(), E = M->end(); I != E;) {
+    auto F = I++;
+    if (F->hasName() && F->isDeclaration()) {
+      DEBUG(dbgs() << "[postProcessOCL sret] " << *F << '\n');
+      if (F->getReturnType()->isStructTy() &&
+          oclIsBuiltin(F->getName(), &DemangledName, isCPP)) {
+        if (!postProcessOCLBuiltinReturnStruct(&*F))
+          return false;
+      }
+    }
+  }
+  for (auto I = M->begin(), E = M->end(); I != E;) {
+    auto F = I++;
+    if (F->hasName() && F->isDeclaration()) {
+      DEBUG(dbgs() << "[postProcessOCL func ptr] " << *F << '\n');
+      auto AI = F->arg_begin();
+      if (hasFunctionPointerArg(&*F, AI) && isDecoratedSPIRVFunc(&*F))
+        if (!postProcessOCLBuiltinWithFuncPointer(&*F, AI))
+          return false;
+    }
+  }
+  for (auto I = M->begin(), E = M->end(); I != E;) {
+    auto F = I++;
+    if (F->hasName() && F->isDeclaration()) {
+      DEBUG(dbgs() << "[postProcessOCL array arg] " << *F << '\n');
+      if (hasArrayArg(&*F) && oclIsBuiltin(F->getName(), &DemangledName, isCPP))
+        if (!postProcessOCLBuiltinWithArrayArguments(&*F, DemangledName))
+          return false;
+    }
+  }
+  return true;
+}
+
+bool
+SPIRVToLLVM::postProcessOCLBuiltinReturnStruct(Function *F) {
+  std::string Name = F->getName();
+  F->setName(Name + ".old");
+  for (auto I = F->user_begin(), E = F->user_end(); I != E;) {
+    if (auto CI = dyn_cast<CallInst>(*I++)) {
+      auto ST = dyn_cast<StoreInst>(*(CI->user_begin()));
       assert(ST);
-      std::vector<Type *> ArgTys;
-      getFunctionTypeParameterTypes(F->getFunctionType(), ArgTys);
-      ArgTys.insert(ArgTys.begin(), PointerType::get(F->getReturnType(),
-          SPIRAS_Private));
-      auto newF = getOrCreateFunction(M, Type::getVoidTy(*Context),
-          ArgTys, Name);
-      newF->setCallingConv(F->getCallingConv());
-      auto Args = getArguments(CI);
-      Args.insert(Args.begin(), ST->getPointerOperand());
-      auto NewCI = CallInst::Create(newF, Args, CI->getName(), CI);
-      NewCI->setCallingConv(CI->getCallingConv());
-      ST->dropAllReferences();
-      ST->removeFromParent();
-      CI->dropAllReferences();
-      CI->removeFromParent();
-    }
-  }
-  F->dropAllReferences();
-  F->removeFromParent();
-  return true;
-}
-
-bool
-SPIRVToLLVM::postProcessOCLBuiltinWithFuncPointer(Function* F,
-    Function::arg_iterator I) {
-  auto Name = undecorateSPIRVFunction(F->getName());
-  std::set<Value *> InvokeFuncPtrs;
-  mutateFunctionOCL (F, [=, &InvokeFuncPtrs](
-      CallInst *CI, std::vector<Value *> &Args) {
-    auto ALoc = std::find_if(Args.begin(), Args.end(), [](Value * elem) {
-        return isFunctionPointerType(elem->getType());
-      });
-    assert(ALoc != Args.end() && "Buit-in must accept a pointer to function");
-    assert(isa<Function>(*ALoc) && "Invalid function pointer usage");
-    Value *Ctx = ALoc[1];
-    Value *CtxLen = ALoc[2];
-    Value *CtxAlign = ALoc[3];
-    if (Name == kOCLBuiltinName::EnqueueKernel)
-      assert(Args.end() - ALoc > 3);
-    else
-      assert(Args.end() - ALoc > 0);
-    // Erase arguments what are hanled by "spir_block_bind" according to SPIR 2.0
-    Args.erase(ALoc + 1, ALoc + 4);
-
-    InvokeFuncPtrs.insert(*ALoc);
-    // There will be as many calls to spir_block_bind as how much device execution
-    // bult-ins using this block. This doesn't contradict SPIR 2.0 specification.
-    *ALoc = addBlockBind(M, cast<Function>(removeCast(*ALoc)),
-        Ctx, CtxLen, CtxAlign, CI);
-    return Name;
-  });
-  for (auto &I:InvokeFuncPtrs)
-    eraseIfNoUse(I);
-  return true;
-}
-
-bool
-SPIRVToLLVM::postProcessOCLBuiltinWithArrayArguments(Function* F,
-    const std::string &DemangledName) {
-  DEBUG(dbgs() << "[postProcessOCLBuiltinWithArrayArguments] " << *F << '\n');
-  auto Attrs = F->getAttributes();
-  auto Name = F->getName();
-  mutateFunction(F, [=](CallInst *CI, std::vector<Value *> &Args) {
-    auto FBegin = CI->getParent()->getParent()->begin()->getFirstInsertionPt();
-    for (auto &I:Args) {
-      auto T = I->getType();
-      if (!T->isArrayTy())
-        continue;
-      auto Alloca = new AllocaInst(T, "", FBegin);
-      auto Store = new StoreInst(I, Alloca, false, CI);
-      auto Zero = ConstantInt::getNullValue(Type::getInt32Ty(T->getContext()));
-      Value *Index[] = {Zero, Zero};
-      I = GetElementPtrInst::CreateInBounds(Alloca, Index, "", CI);
-    }
-    return Name;
-  }, nullptr, &Attrs);
-  return true;
-}
-
-// ToDo: Handle unsigned integer return type. May need spec change.
-Instruction *
-SPIRVToLLVM::postProcessOCLReadImage(SPIRVInstruction *BI, CallInst* CI,
-    const std::string &FuncName) {
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  StringRef ImageTypeName;
-  bool isDepthImage = false;
-  if (isOCLImageType(
-          (cast<CallInst>(CI->getOperand(0)))->getArgOperand(0)->getType(),
-          &ImageTypeName))
-    isDepthImage = ImageTypeName.endswith("depth_t");
-  return mutateCallInstOCL(
-      M, CI,
-      [=](CallInst *, std::vector<Value *> &Args, llvm::Type *&RetTy) {
-        CallInst *CallSampledImg = cast<CallInst>(Args[0]);
-        auto Img = CallSampledImg->getArgOperand(0);
-        assert(isOCLImageType(Img->getType()));
-        auto Sampler = CallSampledImg->getArgOperand(1);
-        Args[0] = Img;
-        Args.insert(Args.begin() + 1, Sampler);
-        if(Args.size() > 4 ) {
-          ConstantInt* ImOp = dyn_cast<ConstantInt>(Args[3]);
-          ConstantFP* LodVal = dyn_cast<ConstantFP>(Args[4]);
-          // Drop "Image Operands" argument.
-          Args.erase(Args.begin() + 3, Args.begin() + 4);
-          // If the image operand is LOD and its value is zero, drop it too.
-          if (ImOp && LodVal && LodVal->isNullValue() &&
-              ImOp->getZExtValue() == ImageOperandsMask::ImageOperandsLodMask )
-            Args.erase(Args.begin() + 3, Args.end());
-        }
-        if (CallSampledImg->hasOneUse()) {
-          CallSampledImg->replaceAllUsesWith(
-              UndefValue::get(CallSampledImg->getType()));
-          CallSampledImg->dropAllReferences();
-          CallSampledImg->eraseFromParent();
-        }
-        Type *T = CI->getType();
-        if (auto VT = dyn_cast<VectorType>(T))
-          T = VT->getElementType();
-        RetTy = isDepthImage ? T : CI->getType();
-        return std::string(kOCLBuiltinName::SampledReadImage) +
-               (T->isFloatingPointTy() ? 'f' : 'i');
-      },
-      [=](CallInst *NewCI) -> Instruction * {
-        if (isDepthImage)
-          return InsertElementInst::Create(
-              UndefValue::get(VectorType::get(NewCI->getType(), 4)), NewCI,
-              getSizet(M, 0), "", NewCI->getParent());
-        return NewCI;
-      },
-      &Attrs);
-}
-
-CallInst*
-SPIRVToLLVM::postProcessOCLWriteImage(SPIRVInstruction *BI, CallInst *CI,
-                                      const std::string &DemangledName) {
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  return mutateCallInstOCL(M, CI, [=](CallInst *, std::vector<Value *> &Args) {
-    llvm::Type *T = Args[2]->getType();
-    if (Args.size() > 4) {
-      ConstantInt* ImOp = dyn_cast<ConstantInt>(Args[3]);
-      ConstantFP* LodVal = dyn_cast<ConstantFP>(Args[4]);
-      // Drop "Image Operands" argument.
-      Args.erase(Args.begin() + 3, Args.begin() + 4);
-      // If the image operand is LOD and its value is zero, drop it too.
-      if (ImOp && LodVal && LodVal->isNullValue() &&
-          ImOp->getZExtValue() == ImageOperandsMask::ImageOperandsLodMask )
-        Args.erase(Args.begin() + 3, Args.end());
-      else
-        std::swap(Args[2], Args[3]);
-    }
-    return std::string(kOCLBuiltinName::WriteImage) +
-            (T->isFPOrFPVectorTy() ? 'f' : 'i');
-    }, &Attrs);
-}
-
-CallInst *
-SPIRVToLLVM::postProcessOCLBuildNDRange(SPIRVInstruction *BI, CallInst *CI,
-    const std::string &FuncName) {
-  assert(CI->getNumArgOperands() == 3);
-  auto GWS = CI->getArgOperand(0);
-  auto LWS = CI->getArgOperand(1);
-  auto GWO = CI->getArgOperand(2);
-  CI->setArgOperand(0, GWO);
-  CI->setArgOperand(1, GWS);
-  CI->setArgOperand(2, LWS);
-  return CI;
-}
-
-Instruction *
-SPIRVToLLVM::postProcessGroupAllAny(CallInst *CI,
-                                    const std::string &DemangledName) {
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  return mutateCallInstSPIRV(
-      M, CI,
-      [=](CallInst *, std::vector<Value *> &Args, llvm::Type *&RetTy) {
-        Type *Int32Ty = Type::getInt32Ty(*Context);
-        RetTy = Int32Ty;
-        Args[1] = CastInst::CreateZExtOrBitCast(Args[1], Int32Ty, "", CI);
-        return DemangledName;
-      },
-      [=](CallInst *NewCI) -> Instruction * {
-        Type *RetTy = Type::getInt1Ty(*Context);
-        return CastInst::CreateTruncOrBitCast(NewCI, RetTy, "",
-                                              NewCI->getNextNode());
-      },
-      &Attrs);
-}
-
-CallInst *
-SPIRVToLLVM::expandOCLBuiltinWithScalarArg(CallInst* CI,
-    const std::string &FuncName) {
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  if (!CI->getOperand(0)->getType()->isVectorTy() &&
-    CI->getOperand(1)->getType()->isVectorTy()) {
-    return mutateCallInstOCL(M, CI, [=](CallInst *, std::vector<Value *> &Args){
-      unsigned vecSize = CI->getOperand(1)->getType()->getVectorNumElements();
-      Value *NewVec = nullptr;
-      if (auto CA = dyn_cast<Constant>(Args[0]))
-        NewVec = ConstantVector::getSplat(vecSize, CA);
-      else {
-        NewVec = ConstantVector::getSplat(vecSize,
-            Constant::getNullValue(Args[0]->getType()));
-        NewVec = InsertElementInst::Create(NewVec, Args[0], getInt32(M, 0), "",
-            CI);
-        NewVec = new ShuffleVectorInst(NewVec, NewVec,
-            ConstantVector::getSplat(vecSize, getInt32(M, 0)), "", CI);
-      }
-      NewVec->takeName(Args[0]);
-      Args[0] = NewVec;
-      return FuncName;
-    }, &Attrs);
-  }
-  return CI;
-}
-
-std::string
-SPIRVToLLVM::transOCLPipeTypeAccessQualifier(SPIRV::SPIRVTypePipe* ST) {
-  return SPIRSPIRVAccessQualifierMap::rmap(ST->getAccessQualifier());
-}
-
-void
-SPIRVToLLVM::transGeneratorMD() {
-  SPIRVMDBuilder B(*M);
-  B.addNamedMD(kSPIRVMD::Generator)
-      .addOp()
-        .addU16(BM->getGeneratorId())
-        .addU16(BM->getGeneratorVer())
-        .done();
-}
-
-Value *
-SPIRVToLLVM::oclTransConstantSampler(SPIRV::SPIRVConstantSampler* BCS) {
-  auto Lit = (BCS->getAddrMode() << 1) |
-      BCS->getNormalized() |
-      ((BCS->getFilterMode() + 1) << 4);
-  auto Ty = IntegerType::getInt32Ty(*Context);
-  return ConstantInt::get(Ty, Lit);
-}
-
-/// For instructions, this function assumes they are created in order
-/// and appended to the given basic block. An instruction may use a
-/// instruction from another BB which has not been translated. Such
-/// instructions should be translated to place holders at the point
-/// of first use, then replaced by real instructions when they are
-/// created.
-///
-/// When CreatePlaceHolder is true, create a load instruction of a
-/// global variable as placeholder for SPIRV instruction. Otherwise,
-/// create instruction and replace placeholder if there is one.
-Value *
-SPIRVToLLVM::transValueWithoutDecoration(SPIRVValue *BV, Function *F,
-    BasicBlock *BB, bool CreatePlaceHolder){
-
-  auto OC = BV->getOpCode();
-  IntBoolOpMap::rfind(OC, &OC);
-
-  // Translation of non-instruction values
-  switch(OC) {
-  case OpConstant: {
-    SPIRVConstant *BConst = static_cast<SPIRVConstant *>(BV);
-    SPIRVType *BT = BV->getType();
-    Type *LT = transType(BT);
-    switch(BT->getOpCode()) {
-    case OpTypeBool:
-    case OpTypeInt:
-      return mapValue(BV, ConstantInt::get(LT, BConst->getZExtIntValue(),
-          static_cast<SPIRVTypeInt*>(BT)->isSigned()));
-    case OpTypeFloat: {
-      const llvm::fltSemantics *FS = nullptr;
-      switch (BT->getFloatBitWidth()) {
-      case 16:
-        FS = &APFloat::IEEEhalf;
-        break;
-      case 32:
-        FS = &APFloat::IEEEsingle;
-        break;
-      case 64:
-        FS = &APFloat::IEEEdouble;
-        break;
-      default:
-        llvm_unreachable("invalid float type");
-      }
-      return mapValue(BV, ConstantFP::get(*Context, APFloat(*FS,
-          APInt(BT->getFloatBitWidth(), BConst->getZExtIntValue()))));
-    }
-    default:
-      llvm_unreachable("Not implemented");
-      return nullptr;
-    }
-  }
-
-  case OpConstantTrue:
-    return mapValue(BV, ConstantInt::getTrue(*Context));
-
-  case OpConstantFalse:
-    return mapValue(BV, ConstantInt::getFalse(*Context));
-
-  case OpConstantNull: {
-    auto LT = transType(BV->getType());
-    return mapValue(BV, Constant::getNullValue(LT));
-  }
-
-  case OpConstantComposite: {
-    auto BCC = static_cast<SPIRVConstantComposite*>(BV);
-    std::vector<Constant *> CV;
-    for (auto &I:BCC->getElements())
-      CV.push_back(dyn_cast<Constant>(transValue(I, F, BB)));
-    switch(BV->getType()->getOpCode()) {
-    case OpTypeVector:
-      return mapValue(BV, ConstantVector::get(CV));
-    case OpTypeArray:
-      return mapValue(BV, ConstantArray::get(
-          dyn_cast<ArrayType>(transType(BCC->getType())), CV));
-    case OpTypeStruct:
-      return mapValue(BV, ConstantStruct::get(
-          dyn_cast<StructType>(transType(BCC->getType())), CV));
-    default:
-      llvm_unreachable("not implemented");
-      return nullptr;
-    }
-  }
-
-  case OpConstantSampler: {
-    auto BCS = static_cast<SPIRVConstantSampler*>(BV);
-    return mapValue(BV, oclTransConstantSampler(BCS));
-  }
-
-  case OpSpecConstantOp: {
-    auto BI = createInstFromSpecConstantOp(
-        static_cast<SPIRVSpecConstantOp*>(BV));
-    return mapValue(BV, transValue(BI, nullptr, nullptr, false));
-  }
-
-  case OpUndef:
-    return mapValue(BV, UndefValue::get(transType(BV->getType())));
-
-  case OpVariable: {
-    auto BVar = static_cast<SPIRVVariable *>(BV);
-    auto Ty = transType(BVar->getType()->getPointerElementType());
-    bool IsConst = BVar->isConstant();
-    llvm::GlobalValue::LinkageTypes LinkageTy = transLinkageType(BVar);
-    Constant *Initializer = nullptr;
-    SPIRVValue *Init = BVar->getInitializer();
-    if (Init)
-        Initializer = dyn_cast<Constant>(transValue(Init, F, BB, false));
-    else if (LinkageTy == GlobalValue::CommonLinkage)
-        // In LLVM variables with common linkage type must be initilized by 0
-        Initializer = Constant::getNullValue(Ty);
-
-    SPIRVStorageClassKind BS = BVar->getStorageClass();
-    if (BS == StorageClassFunction && !Init) {
-        assert (BB && "Invalid BB");
-        return mapValue(BV, new AllocaInst(Ty, BV->getName(), BB));
-    }
-    auto AddrSpace = SPIRSPIRVAddrSpaceMap::rmap(BS);
-    auto LVar = new GlobalVariable(*M, Ty, IsConst, LinkageTy, Initializer,
-        BV->getName(), 0, GlobalVariable::NotThreadLocal, AddrSpace);
-    LVar->setUnnamedAddr(IsConst && Ty->isArrayTy() &&
-        Ty->getArrayElementType()->isIntegerTy(8));
-    SPIRVBuiltinVariableKind BVKind;
-    if (BVar->isBuiltin(&BVKind))
-      BuiltinGVMap[LVar] = BVKind;
-    return mapValue(BV, LVar);
-  }
-
-  case OpFunctionParameter: {
-    auto BA = static_cast<SPIRVFunctionParameter*>(BV);
-    assert (F && "Invalid function");
-    unsigned ArgNo = 0;
-    for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
-        ++I, ++ArgNo) {
-      if (ArgNo == BA->getArgNo())
-        return mapValue(BV, I);
-    }
-    llvm_unreachable("Invalid argument");
-    return nullptr;
-  }
-
-  case OpFunction:
-    return mapValue(BV, transFunction(static_cast<SPIRVFunction *>(BV)));
-
-  case OpLabel:
-    return mapValue(BV, BasicBlock::Create(*Context, BV->getName(), F));
-
-  case OpBitcast: // Can be translated without BB pointer
-    if(!CreatePlaceHolder) // May be a placeholder
-      return mapValue(BV, transConvertInst(BV, F, BB));
-
-  default:
-    // do nothing
-    break;
-  }
-
-  // All other values require valid BB pointer.
-  assert(BB && "Invalid BB");
-
-  // Creation of place holder
-  if (CreatePlaceHolder) {
-    auto GV = new GlobalVariable(*M,
-        transType(BV->getType()),
-        false,
-        GlobalValue::PrivateLinkage,
-        nullptr,
-        std::string(kPlaceholderPrefix) + BV->getName(),
-        0, GlobalVariable::NotThreadLocal, 0);
-    auto LD = new LoadInst(GV, BV->getName(), BB);
-    PlaceholderMap[BV] = LD;
-    return mapValue(BV, LD);
-  }
-
-  // Translation of instructions
-  switch (BV->getOpCode()) {
-  case OpBranch: {
-    auto BR = static_cast<SPIRVBranch *>(BV);
-    return mapValue(BV, BranchInst::Create(
-      dyn_cast<BasicBlock>(transValue(BR->getTargetLabel(), F, BB)), BB));
-  }
-
-  case OpBranchConditional: {
-    auto BR = static_cast<SPIRVBranchConditional *>(BV);
-    return mapValue(
-        BV, BranchInst::Create(
-                dyn_cast<BasicBlock>(transValue(BR->getTrueLabel(), F, BB)),
-                dyn_cast<BasicBlock>(transValue(BR->getFalseLabel(), F, BB)),
-                transValue(BR->getCondition(), F, BB), BB));
-  }
-
-  case OpPhi: {
-    auto Phi = static_cast<SPIRVPhi *>(BV);
-    auto LPhi = dyn_cast<PHINode>(mapValue(
-        BV, PHINode::Create(transType(Phi->getType()),
-                            Phi->getPairs().size() / 2, Phi->getName(), BB)));
-    Phi->foreachPair([&](SPIRVValue *IncomingV, SPIRVBasicBlock *IncomingBB,
-                         size_t Index) {
-      auto Translated = transValue(IncomingV, F, BB);
-      LPhi->addIncoming(Translated,
-                        dyn_cast<BasicBlock>(transValue(IncomingBB, F, BB)));
-    });
-    return LPhi;
-  }
-
-  case OpReturn:
-    return mapValue(BV, ReturnInst::Create(*Context, BB));
-
-  case OpReturnValue: {
-    auto RV = static_cast<SPIRVReturnValue *>(BV);
-    return mapValue(
-        BV, ReturnInst::Create(*Context,
-                               transValue(RV->getReturnValue(), F, BB), BB));
-  }
-
-  case OpStore: {
-    SPIRVStore *BS = static_cast<SPIRVStore*>(BV);
-    return mapValue(BV,
-                    new StoreInst(transValue(BS->getSrc(), F, BB),
-                                  transValue(BS->getDst(), F, BB),
-                                  BS->SPIRVMemoryAccess::isVolatile(),
-                                  BS->SPIRVMemoryAccess::getAlignment(), BB));
-  }
-
-  case OpLoad: {
-    SPIRVLoad *BL = static_cast<SPIRVLoad*>(BV);
-    return mapValue(BV,
-                    new LoadInst(transValue(BL->getSrc(), F, BB), BV->getName(),
-                                 BL->SPIRVMemoryAccess::isVolatile(),
-                                 BL->SPIRVMemoryAccess::getAlignment(), BB));
-  }
-
-  case OpCopyMemorySized: {
-    SPIRVCopyMemorySized *BC = static_cast<SPIRVCopyMemorySized *>(BV);
-    std::string FuncName = "llvm.memcpy";
-    SPIRVType* BS = BC->getSource()->getType();
-    SPIRVType* BT = BC->getTarget()->getType();
-    Type *Int1Ty = Type::getInt1Ty(*Context);
-    Type* Int32Ty = Type::getInt32Ty(*Context);
-    Type* VoidTy = Type::getVoidTy(*Context);
-    Type* SrcTy = transType(BS);
-    Type* TrgTy = transType(BT);
-    Type* SizeTy = transType(BC->getSize()->getType());
-    Type* ArgTy[] = { TrgTy, SrcTy, SizeTy, Int32Ty, Int1Ty };
-
-    ostringstream TempName;
-    TempName << ".p" << SPIRSPIRVAddrSpaceMap::rmap(BT->getPointerStorageClass()) << "i8";
-    TempName << ".p" << SPIRSPIRVAddrSpaceMap::rmap(BS->getPointerStorageClass()) << "i8";
-    FuncName += TempName.str();
-    if (BC->getSize()->getType()->getBitWidth() == 32)
-      FuncName += ".i32";
-    else
-      FuncName += ".i64";
-
-    FunctionType *FT = FunctionType::get(VoidTy, ArgTy, false);
-    Function *Func = dyn_cast<Function>(M->getOrInsertFunction(FuncName, FT));
-    assert(Func && Func->getFunctionType() == FT && "Function type mismatch");
-    Func->setLinkage(GlobalValue::ExternalLinkage);
-
-    if (isFuncNoUnwind())
-      Func->addFnAttr(Attribute::NoUnwind);
-
-    Value *Arg[] = { transValue(BC->getTarget(), Func, BB),
-                     transValue(BC->getSource(), Func, BB),
-                     dyn_cast<llvm::ConstantInt>(transValue(BC->getSize(),
-                         Func, BB)),
-                     ConstantInt::get(Int32Ty,
-                         BC->SPIRVMemoryAccess::getAlignment()),
-                     ConstantInt::get(Int1Ty,
-                         BC->SPIRVMemoryAccess::isVolatile())};
-    return mapValue( BV, CallInst::Create(Func, Arg, "", BB));
-  }
-
-  case OpSelect: {
-    SPIRVSelect *BS = static_cast<SPIRVSelect*>(BV);
-    return mapValue(BV,
-                    SelectInst::Create(transValue(BS->getCondition(), F, BB),
-                                       transValue(BS->getTrueValue(), F, BB),
-                                       transValue(BS->getFalseValue(), F, BB),
-                                       BV->getName(), BB));
-  }
-
-  case OpSwitch: {
-    auto BS = static_cast<SPIRVSwitch *>(BV);
-    auto Select = transValue(BS->getSelect(), F, BB);
-    auto LS = SwitchInst::Create(
-        Select, dyn_cast<BasicBlock>(transValue(BS->getDefault(), F, BB)),
-        BS->getNumPairs(), BB);
-    BS->foreachPair(
-        [&](SPIRVWord Literal, SPIRVBasicBlock *Label, size_t Index) {
-          LS->addCase(ConstantInt::get(dyn_cast<IntegerType>(Select->getType()),
-                                       Literal),
-                      dyn_cast<BasicBlock>(transValue(Label, F, BB)));
-        });
-    return mapValue(BV, LS);
-  }
-
-  case OpAccessChain:
-  case OpInBoundsAccessChain:
-  case OpPtrAccessChain:
-  case OpInBoundsPtrAccessChain: {
-    auto AC = static_cast<SPIRVAccessChainBase *>(BV);
-    auto Base = transValue(AC->getBase(), F, BB);
-    auto Index = transValue(AC->getIndices(), F, BB);
-    if (!AC->hasPtrIndex())
-      Index.insert(Index.begin(), getInt32(M, 0));
-    auto IsInbound = AC->isInBounds();
-    Value *V = nullptr;
-    if (BB) {
-      auto GEP = GetElementPtrInst::Create(Base, Index, BV->getName(), BB);
-      GEP->setIsInBounds(IsInbound);
-      V = GEP;
-    } else {
-      V = ConstantExpr::getGetElementPtr(dyn_cast<Constant>(Base), Index,
-                                         IsInbound);
-    }
-    return mapValue(BV, V);
-  }
-
-  case OpCompositeExtract: {
-    SPIRVCompositeExtract *CE = static_cast<SPIRVCompositeExtract *>(BV);
-    assert(CE->getComposite()->getType()->isTypeVector() && "Invalid type");
-    assert(CE->getIndices().size() == 1 && "Invalid index");
-    return mapValue(
-        BV, ExtractElementInst::Create(
-                transValue(CE->getComposite(), F, BB),
-                ConstantInt::get(*Context, APInt(32, CE->getIndices()[0])),
-                BV->getName(), BB));
-  }
-
-  case OpVectorExtractDynamic: {
-    auto CE = static_cast<SPIRVVectorExtractDynamic *>(BV);
-    return mapValue(
-        BV, ExtractElementInst::Create(transValue(CE->getVector(), F, BB),
-                                       transValue(CE->getIndex(), F, BB),
-                                       BV->getName(), BB));
-    }
-
-  case OpCompositeInsert: {
-    auto CI = static_cast<SPIRVCompositeInsert *>(BV);
-    assert(CI->getComposite()->getType()->isTypeVector() && "Invalid type");
-    assert(CI->getIndices().size() == 1 && "Invalid index");
-    return mapValue(
-        BV, InsertElementInst::Create(
-                transValue(CI->getComposite(), F, BB),
-                transValue(CI->getObject(), F, BB),
-                ConstantInt::get(*Context, APInt(32, CI->getIndices()[0])),
-                BV->getName(), BB));
-  }
-
-  case OpVectorInsertDynamic: {
-    auto CI = static_cast<SPIRVVectorInsertDynamic *>(BV);
-    return mapValue(
-        BV, InsertElementInst::Create(transValue(CI->getVector(), F, BB),
-                                      transValue(CI->getComponent(), F, BB),
-                                      transValue(CI->getIndex(), F, BB),
-                                      BV->getName(), BB));
-  }
-
-  case OpVectorShuffle: {
-    auto VS = static_cast<SPIRVVectorShuffle *>(BV);
-    std::vector<Constant *> Components;
-    IntegerType *Int32Ty = IntegerType::get(*Context, 32);
-    for (auto I : VS->getComponents()) {
-      if (I == static_cast<SPIRVWord>(-1))
-        Components.push_back(UndefValue::get(Int32Ty));
-      else
-        Components.push_back(ConstantInt::get(Int32Ty, I));
-    }
-    return mapValue(BV,
-                    new ShuffleVectorInst(transValue(VS->getVector1(), F, BB),
-                                          transValue(VS->getVector2(), F, BB),
-                                          ConstantVector::get(Components),
-                                          BV->getName(), BB));
-  }
-
-  case OpFunctionCall: {
-    SPIRVFunctionCall *BC = static_cast<SPIRVFunctionCall *>(BV);
-    auto Call = CallInst::Create(transFunction(BC->getFunction()),
-                                 transValue(BC->getArgumentValues(), F, BB),
-                                 BC->getName(), BB);
-    setCallingConv(Call);
-    setAttrByCalledFunc(Call);
-    return mapValue(BV, Call);
-  }
-
-  case OpExtInst:
-    return mapValue(
-        BV, transOCLBuiltinFromExtInst(static_cast<SPIRVExtInst *>(BV), BB));
-
-  case OpControlBarrier:
-  case OpMemoryBarrier:
-    return mapValue(
-        BV, transOCLBarrierFence(static_cast<SPIRVInstruction *>(BV), BB));
-
-  case OpSNegate: {
-    SPIRVUnary *BC = static_cast<SPIRVUnary *>(BV);
-    return mapValue(
-        BV, BinaryOperator::CreateNSWNeg(transValue(BC->getOperand(0), F, BB),
-                                         BV->getName(), BB));
-  }
-
-  case OpFNegate: {
-    SPIRVUnary *BC = static_cast<SPIRVUnary *>(BV);
-    return mapValue(
-        BV, BinaryOperator::CreateFNeg(transValue(BC->getOperand(0), F, BB),
-                                       BV->getName(), BB));
-  }
-
-  case OpNot: {
-    SPIRVUnary *BC = static_cast<SPIRVUnary *>(BV);
-    return mapValue(
-        BV, BinaryOperator::CreateNot(transValue(BC->getOperand(0), F, BB),
-                                      BV->getName(), BB));
-  }
-
-  case OpAll :
-  case OpAny :
-    return mapValue(BV,
-                    transOCLAllAny(static_cast<SPIRVInstruction *>(BV), BB));
-
-  case OpIsFinite :
-  case OpIsInf :
-  case OpIsNan :
-  case OpIsNormal :
-  case OpSignBitSet :
-    return mapValue(BV,
-                    transOCLRelational(static_cast<SPIRVInstruction *>(BV), BB));
-
-  default: {
-    auto OC = BV->getOpCode();
-    if (isSPIRVCmpInstTransToLLVMInst(static_cast<SPIRVInstruction*>(BV))) {
-      return mapValue(BV, transCmpInst(BV, BB, F));
-    } else if (OCLSPIRVBuiltinMap::rfind(OC, nullptr) &&
-               !isAtomicOpCode(OC) &&
-               !isGroupOpCode(OC) &&
-               !isPipeOpCode(OC)) {
-      return mapValue(BV, transOCLBuiltinFromInst(
-          static_cast<SPIRVInstruction *>(BV), BB));
-    } else if (isBinaryShiftLogicalBitwiseOpCode(OC) ||
-                isLogicalOpCode(OC)) {
-          return mapValue(BV, transShiftLogicalBitwiseInst(BV, BB, F));
-    } else if (isCvtOpCode(OC)) {
-        auto BI = static_cast<SPIRVInstruction *>(BV);
-        Value *Inst = nullptr;
-        if (BI->hasFPRoundingMode() || BI->isSaturatedConversion())
-          Inst = transOCLBuiltinFromInst(BI, BB);
-        else
-          Inst = transConvertInst(BV, F, BB);
-        return mapValue(BV, Inst);
-    }
-    return mapValue(BV, transSPIRVBuiltinFromInst(
-      static_cast<SPIRVInstruction *>(BV), BB));
-  }
-
-  SPIRVDBG(spvdbgs() << "Cannot translate " << *BV << '\n';)
-  llvm_unreachable("Translation of SPIRV instruction not implemented");
-  return NULL;
-  }
-}
-
-template<class SourceTy, class FuncTy>
-bool
-SPIRVToLLVM::foreachFuncCtlMask(SourceTy Source, FuncTy Func) {
-  SPIRVWord FCM = Source->getFuncCtlMask();
-  SPIRSPIRVFuncCtlMaskMap::foreach([&](Attribute::AttrKind Attr,
-      SPIRVFunctionControlMaskKind Mask){
-    if (FCM & Mask)
-      Func(Attr);
-  });
-  return true;
-}
-
-Function *
-SPIRVToLLVM::transFunction(SPIRVFunction *BF) {
-  auto Loc = FuncMap.find(BF);
-  if (Loc != FuncMap.end())
-    return Loc->second;
-
-  auto IsKernel = BM->isEntryPoint(ExecutionModelKernel, BF->getId());
-  auto Linkage = IsKernel ? GlobalValue::ExternalLinkage : transLinkageType(BF);
-  FunctionType *FT = dyn_cast<FunctionType>(transType(BF->getFunctionType()));
-  Function *F = dyn_cast<Function>(mapValue(BF, Function::Create(FT, Linkage,
-      BF->getName(), M)));
+      std::vector<Type *> ArgTys;
+      getFunctionTypeParameterTypes(F->getFunctionType(), ArgTys);
+      ArgTys.insert(ArgTys.begin(), PointerType::get(F->getReturnType(),
+          SPIRAS_Private));
+      auto newF = getOrCreateFunction(M, Type::getVoidTy(*Context),
+          ArgTys, Name);
+      newF->setCallingConv(F->getCallingConv());
+      auto Args = getArguments(CI);
+      Args.insert(Args.begin(), ST->getPointerOperand());
+      auto NewCI = CallInst::Create(newF, Args, CI->getName(), CI);
+      NewCI->setCallingConv(CI->getCallingConv());
+      ST->dropAllReferences();
+      ST->removeFromParent();
+      CI->dropAllReferences();
+      CI->removeFromParent();
+    }
+  }
+  F->dropAllReferences();
+  F->removeFromParent();
+  return true;
+}
+
+bool
+SPIRVToLLVM::postProcessOCLBuiltinWithFuncPointer(Function* F,
+    Function::arg_iterator I) {
+  auto Name = undecorateSPIRVFunction(F->getName());
+  std::set<Value *> InvokeFuncPtrs;
+  mutateFunctionOCL (F, [=, &InvokeFuncPtrs](
+      CallInst *CI, std::vector<Value *> &Args) {
+    auto ALoc = std::find_if(Args.begin(), Args.end(), [](Value * elem) {
+        return isFunctionPointerType(elem->getType());
+      });
+    assert(ALoc != Args.end() && "Buit-in must accept a pointer to function");
+    assert(isa<Function>(*ALoc) && "Invalid function pointer usage");
+    Value *Ctx = ALoc[1];
+    Value *CtxLen = ALoc[2];
+    Value *CtxAlign = ALoc[3];
+    if (Name == kOCLBuiltinName::EnqueueKernel)
+      assert(Args.end() - ALoc > 3);
+    else
+      assert(Args.end() - ALoc > 0);
+    // Erase arguments what are hanled by "spir_block_bind" according to SPIR 2.0
+    Args.erase(ALoc + 1, ALoc + 4);
+
+    InvokeFuncPtrs.insert(*ALoc);
+    // There will be as many calls to spir_block_bind as how much device execution
+    // bult-ins using this block. This doesn't contradict SPIR 2.0 specification.
+    *ALoc = addBlockBind(M, cast<Function>(removeCast(*ALoc)),
+        Ctx, CtxLen, CtxAlign, CI);
+    return Name;
+  });
+  for (auto &I:InvokeFuncPtrs)
+    eraseIfNoUse(I);
+  return true;
+}
+
+bool
+SPIRVToLLVM::postProcessOCLBuiltinWithArrayArguments(Function* F,
+    const std::string &DemangledName) {
+  DEBUG(dbgs() << "[postProcessOCLBuiltinWithArrayArguments] " << *F << '\n');
+  auto Attrs = F->getAttributes();
+  auto Name = F->getName();
+  mutateFunction(F, [=](CallInst *CI, std::vector<Value *> &Args) {
+    auto FBegin = CI->getParent()->getParent()->begin()->getFirstInsertionPt();
+    for (auto &I:Args) {
+      auto T = I->getType();
+      if (!T->isArrayTy())
+        continue;
+      auto Alloca = new AllocaInst(T, "", &*FBegin);
+      auto Store = new StoreInst(I, Alloca, false, CI);
+      auto Zero = ConstantInt::getNullValue(Type::getInt32Ty(T->getContext()));
+      Value *Index[] = {Zero, Zero};
+      I = GetElementPtrInst::CreateInBounds(Alloca, Index, "", CI);
+    }
+    return Name;
+  }, nullptr, &Attrs);
+  return true;
+}
+
+// ToDo: Handle unsigned integer return type. May need spec change.
+Instruction *
+SPIRVToLLVM::postProcessOCLReadImage(SPIRVInstruction *BI, CallInst* CI,
+    const std::string &FuncName) {
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  StringRef ImageTypeName;
+  bool isDepthImage = false;
+  if (isOCLImageType(
+          (cast<CallInst>(CI->getOperand(0)))->getArgOperand(0)->getType(),
+          &ImageTypeName))
+    isDepthImage = ImageTypeName.endswith("depth_t");
+  return mutateCallInstOCL(
+      M, CI,
+      [=](CallInst *, std::vector<Value *> &Args, llvm::Type *&RetTy) {
+        CallInst *CallSampledImg = cast<CallInst>(Args[0]);
+        auto Img = CallSampledImg->getArgOperand(0);
+        assert(isOCLImageType(Img->getType()));
+        auto Sampler = CallSampledImg->getArgOperand(1);
+        Args[0] = Img;
+        Args.insert(Args.begin() + 1, Sampler);
+        if(Args.size() > 4 ) {
+          ConstantInt* ImOp = dyn_cast<ConstantInt>(Args[3]);
+          ConstantFP* LodVal = dyn_cast<ConstantFP>(Args[4]);
+          // Drop "Image Operands" argument.
+          Args.erase(Args.begin() + 3, Args.begin() + 4);
+          // If the image operand is LOD and its value is zero, drop it too.
+          if (ImOp && LodVal && LodVal->isNullValue() &&
+              ImOp->getZExtValue() == ImageOperandsMask::ImageOperandsLodMask )
+            Args.erase(Args.begin() + 3, Args.end());
+        }
+        if (CallSampledImg->hasOneUse()) {
+          CallSampledImg->replaceAllUsesWith(
+              UndefValue::get(CallSampledImg->getType()));
+          CallSampledImg->dropAllReferences();
+          CallSampledImg->eraseFromParent();
+        }
+        Type *T = CI->getType();
+        if (auto VT = dyn_cast<VectorType>(T))
+          T = VT->getElementType();
+        RetTy = isDepthImage ? T : CI->getType();
+        return std::string(kOCLBuiltinName::SampledReadImage) +
+               (T->isFloatingPointTy() ? 'f' : 'i');
+      },
+      [=](CallInst *NewCI) -> Instruction * {
+        if (isDepthImage)
+          return InsertElementInst::Create(
+              UndefValue::get(VectorType::get(NewCI->getType(), 4)), NewCI,
+              getSizet(M, 0), "", NewCI->getParent());
+        return NewCI;
+      },
+      &Attrs);
+}
+
+CallInst*
+SPIRVToLLVM::postProcessOCLWriteImage(SPIRVInstruction *BI, CallInst *CI,
+                                      const std::string &DemangledName) {
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  return mutateCallInstOCL(M, CI, [=](CallInst *, std::vector<Value *> &Args) {
+    llvm::Type *T = Args[2]->getType();
+    if (Args.size() > 4) {
+      ConstantInt* ImOp = dyn_cast<ConstantInt>(Args[3]);
+      ConstantFP* LodVal = dyn_cast<ConstantFP>(Args[4]);
+      // Drop "Image Operands" argument.
+      Args.erase(Args.begin() + 3, Args.begin() + 4);
+      // If the image operand is LOD and its value is zero, drop it too.
+      if (ImOp && LodVal && LodVal->isNullValue() &&
+          ImOp->getZExtValue() == ImageOperandsMask::ImageOperandsLodMask )
+        Args.erase(Args.begin() + 3, Args.end());
+      else
+        std::swap(Args[2], Args[3]);
+    }
+    return std::string(kOCLBuiltinName::WriteImage) +
+            (T->isFPOrFPVectorTy() ? 'f' : 'i');
+    }, &Attrs);
+}
+
+CallInst *
+SPIRVToLLVM::postProcessOCLBuildNDRange(SPIRVInstruction *BI, CallInst *CI,
+    const std::string &FuncName) {
+  assert(CI->getNumArgOperands() == 3);
+  auto GWS = CI->getArgOperand(0);
+  auto LWS = CI->getArgOperand(1);
+  auto GWO = CI->getArgOperand(2);
+  CI->setArgOperand(0, GWO);
+  CI->setArgOperand(1, GWS);
+  CI->setArgOperand(2, LWS);
+  return CI;
+}
+
+Instruction *
+SPIRVToLLVM::postProcessGroupAllAny(CallInst *CI,
+                                    const std::string &DemangledName) {
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  return mutateCallInstSPIRV(
+      M, CI,
+      [=](CallInst *, std::vector<Value *> &Args, llvm::Type *&RetTy) {
+        Type *Int32Ty = Type::getInt32Ty(*Context);
+        RetTy = Int32Ty;
+        Args[1] = CastInst::CreateZExtOrBitCast(Args[1], Int32Ty, "", CI);
+        return DemangledName;
+      },
+      [=](CallInst *NewCI) -> Instruction * {
+        Type *RetTy = Type::getInt1Ty(*Context);
+        return CastInst::CreateTruncOrBitCast(NewCI, RetTy, "",
+                                              NewCI->getNextNode());
+      },
+      &Attrs);
+}
+
+CallInst *
+SPIRVToLLVM::expandOCLBuiltinWithScalarArg(CallInst* CI,
+    const std::string &FuncName) {
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  if (!CI->getOperand(0)->getType()->isVectorTy() &&
+    CI->getOperand(1)->getType()->isVectorTy()) {
+    return mutateCallInstOCL(M, CI, [=](CallInst *, std::vector<Value *> &Args){
+      unsigned vecSize = CI->getOperand(1)->getType()->getVectorNumElements();
+      Value *NewVec = nullptr;
+      if (auto CA = dyn_cast<Constant>(Args[0]))
+        NewVec = ConstantVector::getSplat(vecSize, CA);
+      else {
+        NewVec = ConstantVector::getSplat(vecSize,
+            Constant::getNullValue(Args[0]->getType()));
+        NewVec = InsertElementInst::Create(NewVec, Args[0], getInt32(M, 0), "",
+            CI);
+        NewVec = new ShuffleVectorInst(NewVec, NewVec,
+            ConstantVector::getSplat(vecSize, getInt32(M, 0)), "", CI);
+      }
+      NewVec->takeName(Args[0]);
+      Args[0] = NewVec;
+      return FuncName;
+    }, &Attrs);
+  }
+  return CI;
+}
+
+std::string
+SPIRVToLLVM::transOCLPipeTypeAccessQualifier(SPIRV::SPIRVTypePipe* ST) {
+  return SPIRSPIRVAccessQualifierMap::rmap(ST->getAccessQualifier());
+}
+
+void
+SPIRVToLLVM::transGeneratorMD() {
+  SPIRVMDBuilder B(*M);
+  B.addNamedMD(kSPIRVMD::Generator)
+      .addOp()
+        .addU16(BM->getGeneratorId())
+        .addU16(BM->getGeneratorVer())
+        .done();
+}
+
+Value *
+SPIRVToLLVM::oclTransConstantSampler(SPIRV::SPIRVConstantSampler* BCS) {
+  auto Lit = (BCS->getAddrMode() << 1) |
+      BCS->getNormalized() |
+      ((BCS->getFilterMode() + 1) << 4);
+  auto Ty = IntegerType::getInt32Ty(*Context);
+  return ConstantInt::get(Ty, Lit);
+}
+
+const llvm::fltSemantics & fltSemOf(int width)
+{
+  switch (width) {
+    case 16:
+      return APFloat::IEEEhalf();
+    case 32:
+      return APFloat::IEEEsingle();
+    case 64:
+      return APFloat::IEEEdouble();
+    default:
+      llvm_unreachable("invalid float type");
+    }
+}
+/// For instructions, this function assumes they are created in order
+/// and appended to the given basic block. An instruction may use a
+/// instruction from another BB which has not been translated. Such
+/// instructions should be translated to place holders at the point
+/// of first use, then replaced by real instructions when they are
+/// created.
+///
+/// When CreatePlaceHolder is true, create a load instruction of a
+/// global variable as placeholder for SPIRV instruction. Otherwise,
+/// create instruction and replace placeholder if there is one.
+Value *
+SPIRVToLLVM::transValueWithoutDecoration(SPIRVValue *BV, Function *F,
+    BasicBlock *BB, bool CreatePlaceHolder){
+
+  auto OC = BV->getOpCode();
+  IntBoolOpMap::rfind(OC, &OC);
+
+  // Translation of non-instruction values
+  switch(OC) {
+  case OpConstant: {
+    SPIRVConstant *BConst = static_cast<SPIRVConstant *>(BV);
+    SPIRVType *BT = BV->getType();
+    Type *LT = transType(BT);
+    switch(BT->getOpCode()) {
+    case OpTypeBool:
+    case OpTypeInt:
+      return mapValue(BV, ConstantInt::get(LT, BConst->getZExtIntValue(),
+          static_cast<SPIRVTypeInt*>(BT)->isSigned()));
+    case OpTypeFloat: {
+      const llvm::fltSemantics &FS = fltSemOf(BT->getFloatBitWidth());
+      return mapValue(BV, ConstantFP::get(*Context, APFloat(FS,
+          APInt(BT->getFloatBitWidth(), BConst->getZExtIntValue()))));
+    }
+    default:
+      llvm_unreachable("Not implemented");
+      return nullptr;
+    }
+  }
+
+  case OpConstantTrue:
+    return mapValue(BV, ConstantInt::getTrue(*Context));
+
+  case OpConstantFalse:
+    return mapValue(BV, ConstantInt::getFalse(*Context));
+
+  case OpConstantNull: {
+    auto LT = transType(BV->getType());
+    return mapValue(BV, Constant::getNullValue(LT));
+  }
+
+  case OpConstantComposite: {
+    auto BCC = static_cast<SPIRVConstantComposite*>(BV);
+    std::vector<Constant *> CV;
+    for (auto &I:BCC->getElements())
+      CV.push_back(dyn_cast<Constant>(transValue(I, F, BB)));
+    switch(BV->getType()->getOpCode()) {
+    case OpTypeVector:
+      return mapValue(BV, ConstantVector::get(CV));
+    case OpTypeArray:
+      return mapValue(BV, ConstantArray::get(
+          dyn_cast<ArrayType>(transType(BCC->getType())), CV));
+    case OpTypeStruct:
+      return mapValue(BV, ConstantStruct::get(
+          dyn_cast<StructType>(transType(BCC->getType())), CV));
+    default:
+      llvm_unreachable("not implemented");
+      return nullptr;
+    }
+  }
+
+  case OpConstantSampler: {
+    auto BCS = static_cast<SPIRVConstantSampler*>(BV);
+    return mapValue(BV, oclTransConstantSampler(BCS));
+  }
+
+  case OpSpecConstantOp: {
+    auto BI = createInstFromSpecConstantOp(
+        static_cast<SPIRVSpecConstantOp*>(BV));
+    return mapValue(BV, transValue(BI, nullptr, nullptr, false));
+  }
+
+  case OpUndef:
+    return mapValue(BV, UndefValue::get(transType(BV->getType())));
+
+  case OpVariable: {
+    auto BVar = static_cast<SPIRVVariable *>(BV);
+    auto Ty = transType(BVar->getType()->getPointerElementType());
+    bool IsConst = BVar->isConstant();
+    llvm::GlobalValue::LinkageTypes LinkageTy = transLinkageType(BVar);
+    Constant *Initializer = nullptr;
+    SPIRVValue *Init = BVar->getInitializer();
+    if (Init)
+        Initializer = dyn_cast<Constant>(transValue(Init, F, BB, false));
+    else if (LinkageTy == GlobalValue::CommonLinkage)
+        // In LLVM variables with common linkage type must be initilized by 0
+        Initializer = Constant::getNullValue(Ty);
+
+    SPIRVStorageClassKind BS = BVar->getStorageClass();
+    if (BS == StorageClassFunction && !Init) {
+        assert (BB && "Invalid BB");
+        return mapValue(BV, new AllocaInst(Ty, BV->getName(), BB));
+    }
+    auto AddrSpace = SPIRSPIRVAddrSpaceMap::rmap(BS);
+    auto LVar = new GlobalVariable(*M, Ty, IsConst, LinkageTy, Initializer,
+        BV->getName(), 0, GlobalVariable::NotThreadLocal, AddrSpace);
+    LVar->setUnnamedAddr(IsConst && Ty->isArrayTy() &&
+                         Ty->getArrayElementType()->isIntegerTy(8) ? GlobalValue::UnnamedAddr::Global
+                                                                   : GlobalValue::UnnamedAddr::None);
+    SPIRVBuiltinVariableKind BVKind;
+    if (BVar->isBuiltin(&BVKind))
+      BuiltinGVMap[LVar] = BVKind;
+    return mapValue(BV, LVar);
+  }
+
+  case OpFunctionParameter: {
+    auto BA = static_cast<SPIRVFunctionParameter*>(BV);
+    assert (F && "Invalid function");
+    unsigned ArgNo = 0;
+    for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
+        ++I, ++ArgNo) {
+      if (ArgNo == BA->getArgNo())
+        return mapValue(BV, &*I);
+    }
+    llvm_unreachable("Invalid argument");
+    return nullptr;
+  }
+
+  case OpFunction:
+    return mapValue(BV, transFunction(static_cast<SPIRVFunction *>(BV)));
+
+  case OpLabel:
+    return mapValue(BV, BasicBlock::Create(*Context, BV->getName(), F));
+
+  case OpBitcast: // Can be translated without BB pointer
+    if(!CreatePlaceHolder) // May be a placeholder
+      return mapValue(BV, transConvertInst(BV, F, BB));
+
+  default:
+    // do nothing
+    break;
+  }
+
+  // All other values require valid BB pointer.
+  assert(BB && "Invalid BB");
+
+  // Creation of place holder
+  if (CreatePlaceHolder) {
+    auto GV = new GlobalVariable(*M,
+        transType(BV->getType()),
+        false,
+        GlobalValue::PrivateLinkage,
+        nullptr,
+        std::string(kPlaceholderPrefix) + BV->getName(),
+        0, GlobalVariable::NotThreadLocal, 0);
+    auto LD = new LoadInst(GV, BV->getName(), BB);
+    PlaceholderMap[BV] = LD;
+    return mapValue(BV, LD);
+  }
+
+  // Translation of instructions
+  switch (BV->getOpCode()) {
+  case OpBranch: {
+    auto BR = static_cast<SPIRVBranch *>(BV);
+    return mapValue(BV, BranchInst::Create(
+      dyn_cast<BasicBlock>(transValue(BR->getTargetLabel(), F, BB)), BB));
+  }
+
+  case OpBranchConditional: {
+    auto BR = static_cast<SPIRVBranchConditional *>(BV);
+    return mapValue(
+        BV, BranchInst::Create(
+                dyn_cast<BasicBlock>(transValue(BR->getTrueLabel(), F, BB)),
+                dyn_cast<BasicBlock>(transValue(BR->getFalseLabel(), F, BB)),
+                transValue(BR->getCondition(), F, BB), BB));
+  }
+
+  case OpPhi: {
+    auto Phi = static_cast<SPIRVPhi *>(BV);
+    auto LPhi = dyn_cast<PHINode>(mapValue(
+        BV, PHINode::Create(transType(Phi->getType()),
+                            Phi->getPairs().size() / 2, Phi->getName(), BB)));
+    Phi->foreachPair([&](SPIRVValue *IncomingV, SPIRVBasicBlock *IncomingBB,
+                         size_t Index) {
+      auto Translated = transValue(IncomingV, F, BB);
+      LPhi->addIncoming(Translated,
+                        dyn_cast<BasicBlock>(transValue(IncomingBB, F, BB)));
+    });
+    return LPhi;
+  }
+
+  case OpReturn:
+    return mapValue(BV, ReturnInst::Create(*Context, BB));
+
+  case OpReturnValue: {
+    auto RV = static_cast<SPIRVReturnValue *>(BV);
+    return mapValue(
+        BV, ReturnInst::Create(*Context,
+                               transValue(RV->getReturnValue(), F, BB), BB));
+  }
+
+  case OpStore: {
+    SPIRVStore *BS = static_cast<SPIRVStore*>(BV);
+    return mapValue(BV,
+                    new StoreInst(transValue(BS->getSrc(), F, BB),
+                                  transValue(BS->getDst(), F, BB),
+                                  BS->SPIRVMemoryAccess::isVolatile(),
+                                  BS->SPIRVMemoryAccess::getAlignment(), BB));
+  }
+
+  case OpLoad: {
+    SPIRVLoad *BL = static_cast<SPIRVLoad*>(BV);
+    return mapValue(BV,
+                    new LoadInst(transValue(BL->getSrc(), F, BB), BV->getName(),
+                                 BL->SPIRVMemoryAccess::isVolatile(),
+                                 BL->SPIRVMemoryAccess::getAlignment(), BB));
+  }
+
+  case OpCopyMemorySized: {
+    SPIRVCopyMemorySized *BC = static_cast<SPIRVCopyMemorySized *>(BV);
+    std::string FuncName = "llvm.memcpy";
+    SPIRVType* BS = BC->getSource()->getType();
+    SPIRVType* BT = BC->getTarget()->getType();
+    Type *Int1Ty = Type::getInt1Ty(*Context);
+    Type* Int32Ty = Type::getInt32Ty(*Context);
+    Type* VoidTy = Type::getVoidTy(*Context);
+    Type* SrcTy = transType(BS);
+    Type* TrgTy = transType(BT);
+    Type* SizeTy = transType(BC->getSize()->getType());
+    Type* ArgTy[] = { TrgTy, SrcTy, SizeTy, Int32Ty, Int1Ty };
+
+    ostringstream TempName;
+    TempName << ".p" << SPIRSPIRVAddrSpaceMap::rmap(BT->getPointerStorageClass()) << "i8";
+    TempName << ".p" << SPIRSPIRVAddrSpaceMap::rmap(BS->getPointerStorageClass()) << "i8";
+    FuncName += TempName.str();
+    if (BC->getSize()->getType()->getBitWidth() == 32)
+      FuncName += ".i32";
+    else
+      FuncName += ".i64";
+
+    FunctionType *FT = FunctionType::get(VoidTy, ArgTy, false);
+    Function *Func = dyn_cast<Function>(M->getOrInsertFunction(FuncName, FT));
+    assert(Func && Func->getFunctionType() == FT && "Function type mismatch");
+    Func->setLinkage(GlobalValue::ExternalLinkage);
+
+    if (isFuncNoUnwind())
+      Func->addFnAttr(Attribute::NoUnwind);
+
+    Value *Arg[] = { transValue(BC->getTarget(), Func, BB),
+                     transValue(BC->getSource(), Func, BB),
+                     dyn_cast<llvm::ConstantInt>(transValue(BC->getSize(),
+                         Func, BB)),
+                     ConstantInt::get(Int32Ty,
+                         BC->SPIRVMemoryAccess::getAlignment()),
+                     ConstantInt::get(Int1Ty,
+                         BC->SPIRVMemoryAccess::isVolatile())};
+    return mapValue( BV, CallInst::Create(Func, Arg, "", BB));
+  }
+
+  case OpSelect: {
+    SPIRVSelect *BS = static_cast<SPIRVSelect*>(BV);
+    return mapValue(BV,
+                    SelectInst::Create(transValue(BS->getCondition(), F, BB),
+                                       transValue(BS->getTrueValue(), F, BB),
+                                       transValue(BS->getFalseValue(), F, BB),
+                                       BV->getName(), BB));
+  }
+
+  case OpSwitch: {
+    auto BS = static_cast<SPIRVSwitch *>(BV);
+    auto Select = transValue(BS->getSelect(), F, BB);
+    auto LS = SwitchInst::Create(
+        Select, dyn_cast<BasicBlock>(transValue(BS->getDefault(), F, BB)),
+        BS->getNumPairs(), BB);
+    BS->foreachPair(
+        [&](SPIRVWord Literal, SPIRVBasicBlock *Label, size_t Index) {
+          LS->addCase(ConstantInt::get(dyn_cast<IntegerType>(Select->getType()),
+                                       Literal),
+                      dyn_cast<BasicBlock>(transValue(Label, F, BB)));
+        });
+    return mapValue(BV, LS);
+  }
+
+  case OpAccessChain:
+  case OpInBoundsAccessChain:
+  case OpPtrAccessChain:
+  case OpInBoundsPtrAccessChain: {
+    auto AC = static_cast<SPIRVAccessChainBase *>(BV);
+    auto Base = transValue(AC->getBase(), F, BB);
+    auto Index = transValue(AC->getIndices(), F, BB);
+    if (!AC->hasPtrIndex())
+      Index.insert(Index.begin(), getInt32(M, 0));
+    auto IsInbound = AC->isInBounds();
+    Value *V = nullptr;
+    if (BB) {
+      auto GEP = GetElementPtrInst::Create(Base->getType(),Base, Index, BV->getName(), BB);
+      GEP->setIsInBounds(IsInbound);
+      V = GEP;
+    } else {
+      auto c = dyn_cast<Constant>(Base);
+      V = ConstantExpr::getGetElementPtr(c->getType(), c, Index,
+                                         IsInbound);
+    }
+    return mapValue(BV, V);
+  }
+
+  case OpCompositeExtract: {
+    SPIRVCompositeExtract *CE = static_cast<SPIRVCompositeExtract *>(BV);
+    assert(CE->getComposite()->getType()->isTypeVector() && "Invalid type");
+    assert(CE->getIndices().size() == 1 && "Invalid index");
+    return mapValue(
+        BV, ExtractElementInst::Create(
+                transValue(CE->getComposite(), F, BB),
+                ConstantInt::get(*Context, APInt(32, CE->getIndices()[0])),
+                BV->getName(), BB));
+  }
+
+  case OpVectorExtractDynamic: {
+    auto CE = static_cast<SPIRVVectorExtractDynamic *>(BV);
+    return mapValue(
+        BV, ExtractElementInst::Create(transValue(CE->getVector(), F, BB),
+                                       transValue(CE->getIndex(), F, BB),
+                                       BV->getName(), BB));
+    }
+
+  case OpCompositeInsert: {
+    auto CI = static_cast<SPIRVCompositeInsert *>(BV);
+    assert(CI->getComposite()->getType()->isTypeVector() && "Invalid type");
+    assert(CI->getIndices().size() == 1 && "Invalid index");
+    return mapValue(
+        BV, InsertElementInst::Create(
+                transValue(CI->getComposite(), F, BB),
+                transValue(CI->getObject(), F, BB),
+                ConstantInt::get(*Context, APInt(32, CI->getIndices()[0])),
+                BV->getName(), BB));
+  }
+
+  case OpVectorInsertDynamic: {
+    auto CI = static_cast<SPIRVVectorInsertDynamic *>(BV);
+    return mapValue(
+        BV, InsertElementInst::Create(transValue(CI->getVector(), F, BB),
+                                      transValue(CI->getComponent(), F, BB),
+                                      transValue(CI->getIndex(), F, BB),
+                                      BV->getName(), BB));
+  }
+
+  case OpVectorShuffle: {
+    auto VS = static_cast<SPIRVVectorShuffle *>(BV);
+    std::vector<Constant *> Components;
+    IntegerType *Int32Ty = IntegerType::get(*Context, 32);
+    for (auto I : VS->getComponents()) {
+      if (I == static_cast<SPIRVWord>(-1))
+        Components.push_back(UndefValue::get(Int32Ty));
+      else
+        Components.push_back(ConstantInt::get(Int32Ty, I));
+    }
+    return mapValue(BV,
+                    new ShuffleVectorInst(transValue(VS->getVector1(), F, BB),
+                                          transValue(VS->getVector2(), F, BB),
+                                          ConstantVector::get(Components),
+                                          BV->getName(), BB));
+  }
+
+  case OpFunctionCall: {
+    SPIRVFunctionCall *BC = static_cast<SPIRVFunctionCall *>(BV);
+    auto Call = CallInst::Create(transFunction(BC->getFunction()),
+                                 transValue(BC->getArgumentValues(), F, BB),
+                                 BC->getName(), BB);
+    setCallingConv(Call);
+    setAttrByCalledFunc(Call);
+    return mapValue(BV, Call);
+  }
+
+  case OpExtInst:
+    return mapValue(
+        BV, transOCLBuiltinFromExtInst(static_cast<SPIRVExtInst *>(BV), BB));
+
+  case OpControlBarrier:
+  case OpMemoryBarrier:
+    return mapValue(
+        BV, transOCLBarrierFence(static_cast<SPIRVInstruction *>(BV), BB));
+
+  case OpSNegate: {
+    SPIRVUnary *BC = static_cast<SPIRVUnary *>(BV);
+    return mapValue(
+        BV, BinaryOperator::CreateNSWNeg(transValue(BC->getOperand(0), F, BB),
+                                         BV->getName(), BB));
+  }
+
+  case OpFNegate: {
+    SPIRVUnary *BC = static_cast<SPIRVUnary *>(BV);
+    return mapValue(
+        BV, BinaryOperator::CreateFNeg(transValue(BC->getOperand(0), F, BB),
+                                       BV->getName(), BB));
+  }
+
+  case OpNot: {
+    SPIRVUnary *BC = static_cast<SPIRVUnary *>(BV);
+    return mapValue(
+        BV, BinaryOperator::CreateNot(transValue(BC->getOperand(0), F, BB),
+                                      BV->getName(), BB));
+  }
+
+  case OpAll :
+  case OpAny :
+    return mapValue(BV,
+                    transOCLAllAny(static_cast<SPIRVInstruction *>(BV), BB));
+
+  case OpIsFinite :
+  case OpIsInf :
+  case OpIsNan :
+  case OpIsNormal :
+  case OpSignBitSet :
+    return mapValue(BV,
+                    transOCLRelational(static_cast<SPIRVInstruction *>(BV), BB));
+
+  default: {
+    auto OC = BV->getOpCode();
+    if (isSPIRVCmpInstTransToLLVMInst(static_cast<SPIRVInstruction*>(BV))) {
+      return mapValue(BV, transCmpInst(BV, BB, F));
+    } else if (OCLSPIRVBuiltinMap::rfind(OC, nullptr) &&
+               !isAtomicOpCode(OC) &&
+               !isGroupOpCode(OC) &&
+               !isPipeOpCode(OC)) {
+      return mapValue(BV, transOCLBuiltinFromInst(
+          static_cast<SPIRVInstruction *>(BV), BB));
+    } else if (isBinaryShiftLogicalBitwiseOpCode(OC) ||
+                isLogicalOpCode(OC)) {
+          return mapValue(BV, transShiftLogicalBitwiseInst(BV, BB, F));
+    } else if (isCvtOpCode(OC)) {
+        auto BI = static_cast<SPIRVInstruction *>(BV);
+        Value *Inst = nullptr;
+        if (BI->hasFPRoundingMode() || BI->isSaturatedConversion())
+          Inst = transOCLBuiltinFromInst(BI, BB);
+        else
+          Inst = transConvertInst(BV, F, BB);
+        return mapValue(BV, Inst);
+    }
+    return mapValue(BV, transSPIRVBuiltinFromInst(
+      static_cast<SPIRVInstruction *>(BV), BB));
+  }
+
+  SPIRVDBG(spvdbgs() << "Cannot translate " << *BV << '\n';)
+  llvm_unreachable("Translation of SPIRV instruction not implemented");
+  return NULL;
+  }
+}
+
+template<class SourceTy, class FuncTy>
+bool
+SPIRVToLLVM::foreachFuncCtlMask(SourceTy Source, FuncTy Func) {
+  SPIRVWord FCM = Source->getFuncCtlMask();
+  SPIRSPIRVFuncCtlMaskMap::foreach([&](Attribute::AttrKind Attr,
+      SPIRVFunctionControlMaskKind Mask){
+    if (FCM & Mask)
+      Func(Attr);
+  });
+  return true;
+}
+
+Function *
+SPIRVToLLVM::transFunction(SPIRVFunction *BF) {
+  auto Loc = FuncMap.find(BF);
+  if (Loc != FuncMap.end())
+    return Loc->second;
+
+  auto IsKernel = BM->isEntryPoint(ExecutionModelKernel, BF->getId());
+  auto Linkage = IsKernel ? GlobalValue::ExternalLinkage : transLinkageType(BF);
+  FunctionType *FT = dyn_cast<FunctionType>(transType(BF->getFunctionType()));
+  Function *F = dyn_cast<Function>(mapValue(BF, Function::Create(FT, Linkage,
+      BF->getName(), M)));
   assert(F);
-  mapFunction(BF, F);
-  if (!F->isIntrinsic()) {
-    F->setCallingConv(IsKernel ? CallingConv::SPIR_KERNEL :
-        CallingConv::SPIR_FUNC);
-    if (isFuncNoUnwind())
-      F->addFnAttr(Attribute::NoUnwind);
-    foreachFuncCtlMask(BF, [&](Attribute::AttrKind Attr){
-      F->addFnAttr(Attr);
-    });
-  }
-
-  for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
-      ++I) {
-    auto BA = BF->getArgument(I->getArgNo());
-    mapValue(BA, I);
-    setName(I, BA);
-    BA->foreachAttr([&](SPIRVFuncParamAttrKind Kind){
-      if (Kind == FunctionParameterAttributeNoWrite)
-        return;
-      F->addAttribute(I->getArgNo() + 1, SPIRSPIRVFuncParamAttrMap::rmap(Kind));
-    });
-  }
-  BF->foreachReturnValueAttr([&](SPIRVFuncParamAttrKind Kind){
-    if (Kind == FunctionParameterAttributeNoWrite)
-      return;
-    F->addAttribute(AttributeSet::ReturnIndex,
-        SPIRSPIRVFuncParamAttrMap::rmap(Kind));
-  });
-
-  // Creating all basic blocks before creating instructions.
-  for (size_t I = 0, E = BF->getNumBasicBlock(); I != E; ++I) {
-    transValue(BF->getBasicBlock(I), F, nullptr);
-  }
-
-  for (size_t I = 0, E = BF->getNumBasicBlock(); I != E; ++I) {
-    SPIRVBasicBlock *BBB = BF->getBasicBlock(I);
-    BasicBlock *BB = dyn_cast<BasicBlock>(transValue(BBB, F, nullptr));
-    for (size_t BI = 0, BE = BBB->getNumInst(); BI != BE; ++BI) {
-      SPIRVInstruction *BInst = BBB->getInst(BI);
-      transValue(BInst, F, BB, false);
-    }
-  }
-  return F;
-}
-
-/// LLVM convert builtin functions is translated to two instructions:
-/// y = i32 islessgreater(float x, float z) ->
-///     y = i32 ZExt(bool LessGreater(float x, float z))
-/// When translating back, for simplicity, a trunc instruction is inserted
-/// w = bool LessGreater(float x, float z) ->
-///     w = bool Trunc(i32 islessgreater(float x, float z))
-/// Optimizer should be able to remove the redundant trunc/zext
-void
-SPIRVToLLVM::transOCLBuiltinFromInstPreproc(SPIRVInstruction* BI, Type *&RetTy,
-    std::vector<SPIRVValue *> &Args) {
-  if (!BI->hasType())
-    return;
-  auto BT = BI->getType();
-  auto OC = BI->getOpCode();
-  if (isCmpOpCode(BI->getOpCode())) {
-    if (BT->isTypeBool())
-      RetTy = IntegerType::getInt32Ty(*Context);
-    else if (BT->isTypeVectorBool())
-      RetTy = VectorType::get(IntegerType::get(*Context,
-          Args[0]->getType()->getVectorComponentType()->isTypeFloat(64)?64:32),
-          BT->getVectorComponentCount());
-    else
-       llvm_unreachable("invalid compare instruction");
-  } else if (OC == OpGenericCastToPtrExplicit)
-    Args.pop_back();
-  else if (OC == OpImageRead && Args.size() > 2) {
-    // Drop "Image operands" argument
-    Args.erase(Args.begin() + 2);
-  }
-}
-
-Instruction*
-SPIRVToLLVM::transOCLBuiltinPostproc(SPIRVInstruction* BI,
-    CallInst* CI, BasicBlock* BB, const std::string &DemangledName) {
-  auto OC = BI->getOpCode();
-  if (isCmpOpCode(OC) &&
-      BI->getType()->isTypeVectorOrScalarBool()) {
-    return CastInst::Create(Instruction::Trunc, CI, transType(BI->getType()),
-        "cvt", BB);
-  }
-  if (OC == OpImageSampleExplicitLod)
-    return postProcessOCLReadImage(BI, CI, DemangledName);
-  if (OC == OpImageWrite) {
-    return postProcessOCLWriteImage(BI, CI, DemangledName);
-  }
-  if (OC == OpGenericPtrMemSemantics)
-    return BinaryOperator::CreateShl(CI, getInt32(M, 8), "", BB);
-  if (OC == OpImageQueryFormat)
-    return BinaryOperator::CreateSub(
-        CI, getInt32(M, OCLImageChannelDataTypeOffset), "", BB);
-  if (OC == OpImageQueryOrder)
-    return BinaryOperator::CreateSub(
-        CI, getInt32(M, OCLImageChannelOrderOffset), "", BB);
-  if (OC == OpBuildNDRange)
-    return postProcessOCLBuildNDRange(BI, CI, DemangledName);
-  if (OC == OpGroupAll || OC == OpGroupAny)
-    return postProcessGroupAllAny(CI, DemangledName);
-  if (SPIRVEnableStepExpansion &&
-      (DemangledName == "smoothstep" ||
-       DemangledName == "step"))
-    return expandOCLBuiltinWithScalarArg(CI, DemangledName);
-  return CI;
-}
-
-Instruction *
-SPIRVToLLVM::transBuiltinFromInst(const std::string& FuncName,
-    SPIRVInstruction* BI, BasicBlock* BB) {
-  std::string MangledName;
-  auto Ops = BI->getOperands();
-  Type* RetTy = BI->hasType() ? transType(BI->getType()) :
-      Type::getVoidTy(*Context);
-  transOCLBuiltinFromInstPreproc(BI, RetTy, Ops);
-  std::vector<Type*> ArgTys = transTypeVector(
-      SPIRVInstruction::getOperandTypes(Ops));
-  bool HasFuncPtrArg = false;
-  for (auto& I:ArgTys) {
-    if (isa<FunctionType>(I)) {
-      I = PointerType::get(I, SPIRAS_Private);
-      HasFuncPtrArg = true;
-    }
-  }
-  if (!HasFuncPtrArg)
-    MangleOpenCLBuiltin(FuncName, ArgTys, MangledName);
-  else
-    MangledName = decorateSPIRVFunction(FuncName);
-  Function* Func = M->getFunction(MangledName);
-  FunctionType* FT = FunctionType::get(RetTy, ArgTys, false);
-  // ToDo: Some intermediate functions have duplicate names with
-  // different function types. This is OK if the function name
-  // is used internally and finally translated to unique function
-  // names. However it is better to have a way to differentiate
-  // between intermidiate functions and final functions and make
-  // sure final functions have unique names.
-  SPIRVDBG(
-  if (!HasFuncPtrArg && Func && Func->getFunctionType() != FT) {
-    dbgs() << "Warning: Function name conflict:\n"
-       << *Func << '\n'
-       << " => " << *FT << '\n';
-  }
-  )
-  if (!Func || Func->getFunctionType() != FT) {
-    DEBUG(for (auto& I:ArgTys) {
-      dbgs() << *I << '\n';
-    });
-    Func = Function::Create(FT, GlobalValue::ExternalLinkage, MangledName, M);
-    Func->setCallingConv(CallingConv::SPIR_FUNC);
-    if (isFuncNoUnwind())
-      Func->addFnAttr(Attribute::NoUnwind);
-  }
-  auto Call = CallInst::Create(Func,
-      transValue(Ops, BB->getParent(), BB), "", BB);
-  setName(Call, BI);
-  setAttrByCalledFunc(Call);
-  SPIRVDBG(spvdbgs() << "[transInstToBuiltinCall] " << *BI << " -> "; dbgs() <<
-      *Call << '\n';)
-  Instruction *Inst = Call;
-  Inst = transOCLBuiltinPostproc(BI, Call, BB, FuncName);
-  return Inst;
-}
-
-std::string
-SPIRVToLLVM::getOCLBuiltinName(SPIRVInstruction* BI) {
-  auto OC = BI->getOpCode();
-  if (OC == OpGenericCastToPtrExplicit)
-    return getOCLGenericCastToPtrName(BI);
-  if (isCvtOpCode(OC))
-    return getOCLConvertBuiltinName(BI);
-  if (OC == OpBuildNDRange) {
-    auto NDRangeInst = static_cast<SPIRVBuildNDRange *>(BI);
-    auto EleTy = ((NDRangeInst->getOperands())[0])->getType();
-    int Dim = EleTy->isTypeArray() ? EleTy->getArrayLength() : 1;
-    // cygwin does not have std::to_string
-    ostringstream OS;
-    OS << Dim;
-    assert((EleTy->isTypeInt() && Dim == 1) ||
-        (EleTy->isTypeArray() && Dim >= 2 && Dim <= 3));
-    return std::string(kOCLBuiltinName::NDRangePrefix) + OS.str() + "D";
-  }
-  auto Name = OCLSPIRVBuiltinMap::rmap(OC);
-
-  SPIRVType *T = nullptr;
-  switch(OC) {
-  case OpImageRead:
-    T = BI->getType();
-    break;
-  case OpImageWrite:
-    T = BI->getOperands()[2]->getType();
-    break;
-  default:
-    // do nothing
-    break;
-  }
-  if (T && T->isTypeVector())
-    T = T->getVectorComponentType();
-  if (T)
-    Name += T->isTypeFloat()?'f':'i';
-
-  return Name;
-}
-
-Instruction *
-SPIRVToLLVM::transOCLBuiltinFromInst(SPIRVInstruction *BI, BasicBlock *BB) {
-  assert(BB && "Invalid BB");
-  auto FuncName = getOCLBuiltinName(BI);
-  return transBuiltinFromInst(FuncName, BI, BB);
-}
-
-Instruction *
-SPIRVToLLVM::transSPIRVBuiltinFromInst(SPIRVInstruction *BI, BasicBlock *BB) {
-  assert(BB && "Invalid BB");
-  return transBuiltinFromInst(getSPIRVFuncName(BI->getOpCode()), BI, BB);
-}
-
-bool
-SPIRVToLLVM::translate() {
-  if (!transAddressingModel())
-    return false;
-
-  DbgTran.createCompileUnit();
-  DbgTran.addDbgInfoVersion();
-
-  for (unsigned I = 0, E = BM->getNumVariables(); I != E; ++I) {
-    auto BV = BM->getVariable(I);
-    if (BV->getStorageClass() != StorageClassFunction)
-      transValue(BV, nullptr, nullptr);
-  }
-
-  for (unsigned I = 0, E = BM->getNumFunctions(); I != E; ++I) {
-    transFunction(BM->getFunction(I));
-  }
-  if (!transKernelMetadata())
-    return false;
-  if (!transFPContractMetadata())
-    return false;
-  if (!transSourceLanguage())
-    return false;
-  if (!transSourceExtension())
-    return false;
-  transGeneratorMD();
-  if (!transOCLBuiltinsFromVariables())
-    return false;
-  if (!postProcessOCL())
-    return false;
-  eraseUselessFunctions(M);
-  DbgTran.finalize();
-  return true;
-}
-
-bool
-SPIRVToLLVM::transAddressingModel() {
-  switch (BM->getAddressingModel()) {
-  case AddressingModelPhysical64:
-    M->setTargetTriple(SPIR_TARGETTRIPLE64);
-    M->setDataLayout(SPIR_DATALAYOUT64);
-    break;
-  case AddressingModelPhysical32:
-    M->setTargetTriple(SPIR_TARGETTRIPLE32);
-    M->setDataLayout(SPIR_DATALAYOUT32);
-    break;
-  case AddressingModelLogical:
-    // Do not set target triple and data layout
-    break;
-  default:
-    SPIRVCKRT(0, InvalidAddressingModel, "Actual addressing mode is " +
-        (unsigned)BM->getAddressingModel());
-  }
-  return true;
-}
-
-bool
-SPIRVToLLVM::transDecoration(SPIRVValue *BV, Value *V) {
-  if (!transAlign(BV, V))
-    return false;
-  DbgTran.transDbgInfo(BV, V);
-  return true;
-}
-
-bool
-SPIRVToLLVM::transFPContractMetadata() {
-  bool ContractOff = false;
-  for (unsigned I = 0, E = BM->getNumFunctions(); I != E; ++I) {
-    SPIRVFunction *BF = BM->getFunction(I);
-    if (!isOpenCLKernel(BF))
-      continue;
-    if (BF->getExecutionMode(ExecutionModeContractionOff)) {
-      ContractOff = true;
-      break;
-    }
-  }
-  if (!ContractOff)
-    M->getOrInsertNamedMetadata(kSPIR2MD::FPContract);
-  return true;
-}
-
-std::string SPIRVToLLVM::transOCLImageTypeAccessQualifier(
-    SPIRV::SPIRVTypeImage* ST) {
-  return SPIRSPIRVAccessQualifierMap::rmap(ST->getAccessQualifier());
-}
-
-bool
-SPIRVToLLVM::transKernelMetadata() {
-  NamedMDNode *KernelMDs = M->getOrInsertNamedMetadata(SPIR_MD_KERNELS);
-  for (unsigned I = 0, E = BM->getNumFunctions(); I != E; ++I) {
-    SPIRVFunction *BF = BM->getFunction(I);
-    Function *F = static_cast<Function *>(getTranslatedValue(BF));
-    assert(F && "Invalid translated function");
-    if (F->getCallingConv() != CallingConv::SPIR_KERNEL)
-      continue;
-    std::vector<llvm::Metadata*> KernelMD;
-    KernelMD.push_back(ValueAsMetadata::get(F));
-
-    // Generate metadata for kernel_arg_address_spaces
-    addOCLKernelArgumentMetadata(Context, KernelMD,
-        SPIR_MD_KERNEL_ARG_ADDR_SPACE, BF,
-        [=](SPIRVFunctionParameter *Arg){
-      SPIRVType *ArgTy = Arg->getType();
-      SPIRAddressSpace AS = SPIRAS_Private;
-      if (ArgTy->isTypePointer())
-        AS = SPIRSPIRVAddrSpaceMap::rmap(ArgTy->getPointerStorageClass());
-      else if (ArgTy->isTypeOCLImage() || ArgTy->isTypePipe())
-        AS = SPIRAS_Global;
-      return ConstantAsMetadata::get(
-          ConstantInt::get(Type::getInt32Ty(*Context), AS));
-    });
-    // Generate metadata for kernel_arg_access_qual
-    addOCLKernelArgumentMetadata(Context, KernelMD,
-        SPIR_MD_KERNEL_ARG_ACCESS_QUAL, BF,
-        [=](SPIRVFunctionParameter *Arg){
-      std::string Qual;
-      auto T = Arg->getType();
-      if (T->isTypeOCLImage()) {
-        auto ST = static_cast<SPIRVTypeImage *>(T);
-        Qual = transOCLImageTypeAccessQualifier(ST);
-      } else if (T->isTypePipe()){
-        auto PT = static_cast<SPIRVTypePipe *>(T);
-        Qual = transOCLPipeTypeAccessQualifier(PT);
-      } else
-        Qual = "none";
-      return MDString::get(*Context, Qual);
-    });
-    // Generate metadata for kernel_arg_type
-    addOCLKernelArgumentMetadata(Context, KernelMD,
-        SPIR_MD_KERNEL_ARG_TYPE, BF,
-        [=](SPIRVFunctionParameter *Arg){
-      return transOCLKernelArgTypeName(Arg);
-    });
-    // Generate metadata for kernel_arg_type_qual
-    addOCLKernelArgumentMetadata(Context, KernelMD,
-        SPIR_MD_KERNEL_ARG_TYPE_QUAL, BF,
-        [=](SPIRVFunctionParameter *Arg){
-      std::string Qual;
-      if (Arg->hasDecorate(DecorationVolatile))
-        Qual = kOCLTypeQualifierName::Volatile;
-      Arg->foreachAttr([&](SPIRVFuncParamAttrKind Kind){
-        Qual += Qual.empty() ? "" : " ";
-        switch(Kind){
-        case FunctionParameterAttributeNoAlias:
-          Qual += kOCLTypeQualifierName::Restrict;
-          break;
-        case FunctionParameterAttributeNoWrite:
-          Qual += kOCLTypeQualifierName::Const;
-          break;
-        default:
-          // do nothing.
-          break;
-        }
-      });
-      if (Arg->getType()->isTypePipe()) {
-        Qual += Qual.empty() ? "" : " ";
-        Qual += kOCLTypeQualifierName::Pipe;
-      }
-      return MDString::get(*Context, Qual);
-    });
-    // Generate metadata for kernel_arg_base_type
-    addOCLKernelArgumentMetadata(Context, KernelMD,
-        SPIR_MD_KERNEL_ARG_BASE_TYPE, BF,
-        [=](SPIRVFunctionParameter *Arg){
-      return transOCLKernelArgTypeName(Arg);
-    });
-    // Generate metadata for kernel_arg_name
-    if (SPIRVGenKernelArgNameMD) {
-      bool ArgHasName = true;
-      BF->foreachArgument([&](SPIRVFunctionParameter *Arg){
-        ArgHasName &= !Arg->getName().empty();
-      });
-      if (ArgHasName)
-        addOCLKernelArgumentMetadata(Context, KernelMD,
-            SPIR_MD_KERNEL_ARG_NAME, BF,
-            [=](SPIRVFunctionParameter *Arg){
-          return MDString::get(*Context, Arg->getName());
-        });
-    }
-    // Generate metadata for reqd_work_group_size
-    if (auto EM = BF->getExecutionMode(ExecutionModeLocalSize)) {
-      KernelMD.push_back(getMDNodeStringIntVec(Context,
-          kSPIR2MD::WGSize, EM->getLiterals()));
-    }
-    // Generate metadata for work_group_size_hint
-    if (auto EM = BF->getExecutionMode(ExecutionModeLocalSizeHint)) {
-      KernelMD.push_back(getMDNodeStringIntVec(Context,
-          kSPIR2MD::WGSizeHint, EM->getLiterals()));
-    }
-    // Generate metadata for vec_type_hint
-    if (auto EM = BF->getExecutionMode(ExecutionModeVecTypeHint)) {
-      std::vector<Metadata*> MetadataVec;
-      MetadataVec.push_back(MDString::get(*Context, kSPIR2MD::VecTyHint));
-      Type *VecHintTy = decodeVecTypeHint(*Context, EM->getLiterals()[0]);
+  mapFunction(BF, F);
+  if (!F->isIntrinsic()) {
+    F->setCallingConv(IsKernel ? CallingConv::SPIR_KERNEL :
+        CallingConv::SPIR_FUNC);
+    if (isFuncNoUnwind())
+      F->addFnAttr(Attribute::NoUnwind);
+    foreachFuncCtlMask(BF, [&](Attribute::AttrKind Attr){
+      F->addFnAttr(Attr);
+    });
+  }
+
+  for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
+      ++I) {
+    auto BA = BF->getArgument(I->getArgNo());
+    mapValue(BA, &*I);
+    setName(&*I, BA);
+    BA->foreachAttr([&](SPIRVFuncParamAttrKind Kind){
+      if (Kind == FunctionParameterAttributeNoWrite)
+        return;
+      F->addAttribute(I->getArgNo() + 1, SPIRSPIRVFuncParamAttrMap::rmap(Kind));
+    });
+  }
+  BF->foreachReturnValueAttr([&](SPIRVFuncParamAttrKind Kind){
+    if (Kind == FunctionParameterAttributeNoWrite)
+      return;
+    F->addAttribute(AttributeSet::ReturnIndex,
+        SPIRSPIRVFuncParamAttrMap::rmap(Kind));
+  });
+
+  // Creating all basic blocks before creating instructions.
+  for (size_t I = 0, E = BF->getNumBasicBlock(); I != E; ++I) {
+    transValue(BF->getBasicBlock(I), F, nullptr);
+  }
+
+  for (size_t I = 0, E = BF->getNumBasicBlock(); I != E; ++I) {
+    SPIRVBasicBlock *BBB = BF->getBasicBlock(I);
+    BasicBlock *BB = dyn_cast<BasicBlock>(transValue(BBB, F, nullptr));
+    for (size_t BI = 0, BE = BBB->getNumInst(); BI != BE; ++BI) {
+      SPIRVInstruction *BInst = BBB->getInst(BI);
+      transValue(BInst, F, BB, false);
+    }
+  }
+  return F;
+}
+
+/// LLVM convert builtin functions is translated to two instructions:
+/// y = i32 islessgreater(float x, float z) ->
+///     y = i32 ZExt(bool LessGreater(float x, float z))
+/// When translating back, for simplicity, a trunc instruction is inserted
+/// w = bool LessGreater(float x, float z) ->
+///     w = bool Trunc(i32 islessgreater(float x, float z))
+/// Optimizer should be able to remove the redundant trunc/zext
+void
+SPIRVToLLVM::transOCLBuiltinFromInstPreproc(SPIRVInstruction* BI, Type *&RetTy,
+    std::vector<SPIRVValue *> &Args) {
+  if (!BI->hasType())
+    return;
+  auto BT = BI->getType();
+  auto OC = BI->getOpCode();
+  if (isCmpOpCode(BI->getOpCode())) {
+    if (BT->isTypeBool())
+      RetTy = IntegerType::getInt32Ty(*Context);
+    else if (BT->isTypeVectorBool())
+      RetTy = VectorType::get(IntegerType::get(*Context,
+          Args[0]->getType()->getVectorComponentType()->isTypeFloat(64)?64:32),
+          BT->getVectorComponentCount());
+    else
+       llvm_unreachable("invalid compare instruction");
+  } else if (OC == OpGenericCastToPtrExplicit)
+    Args.pop_back();
+  else if (OC == OpImageRead && Args.size() > 2) {
+    // Drop "Image operands" argument
+    Args.erase(Args.begin() + 2);
+  }
+}
+
+Instruction*
+SPIRVToLLVM::transOCLBuiltinPostproc(SPIRVInstruction* BI,
+    CallInst* CI, BasicBlock* BB, const std::string &DemangledName) {
+  auto OC = BI->getOpCode();
+  if (isCmpOpCode(OC) &&
+      BI->getType()->isTypeVectorOrScalarBool()) {
+    return CastInst::Create(Instruction::Trunc, CI, transType(BI->getType()),
+        "cvt", BB);
+  }
+  if (OC == OpImageSampleExplicitLod)
+    return postProcessOCLReadImage(BI, CI, DemangledName);
+  if (OC == OpImageWrite) {
+    return postProcessOCLWriteImage(BI, CI, DemangledName);
+  }
+  if (OC == OpGenericPtrMemSemantics)
+    return BinaryOperator::CreateShl(CI, getInt32(M, 8), "", BB);
+  if (OC == OpImageQueryFormat)
+    return BinaryOperator::CreateSub(
+        CI, getInt32(M, OCLImageChannelDataTypeOffset), "", BB);
+  if (OC == OpImageQueryOrder)
+    return BinaryOperator::CreateSub(
+        CI, getInt32(M, OCLImageChannelOrderOffset), "", BB);
+  if (OC == OpBuildNDRange)
+    return postProcessOCLBuildNDRange(BI, CI, DemangledName);
+  if (OC == OpGroupAll || OC == OpGroupAny)
+    return postProcessGroupAllAny(CI, DemangledName);
+  if (SPIRVEnableStepExpansion &&
+      (DemangledName == "smoothstep" ||
+       DemangledName == "step"))
+    return expandOCLBuiltinWithScalarArg(CI, DemangledName);
+  return CI;
+}
+
+Instruction *
+SPIRVToLLVM::transBuiltinFromInst(const std::string& FuncName,
+    SPIRVInstruction* BI, BasicBlock* BB) {
+  std::string MangledName;
+  auto Ops = BI->getOperands();
+  Type* RetTy = BI->hasType() ? transType(BI->getType()) :
+      Type::getVoidTy(*Context);
+  transOCLBuiltinFromInstPreproc(BI, RetTy, Ops);
+  std::vector<Type*> ArgTys = transTypeVector(
+      SPIRVInstruction::getOperandTypes(Ops));
+  bool HasFuncPtrArg = false;
+  for (auto& I:ArgTys) {
+    if (isa<FunctionType>(I)) {
+      I = PointerType::get(I, SPIRAS_Private);
+      HasFuncPtrArg = true;
+    }
+  }
+  if (!HasFuncPtrArg)
+    MangleOpenCLBuiltin(FuncName, ArgTys, MangledName);
+  else
+    MangledName = decorateSPIRVFunction(FuncName);
+  Function* Func = M->getFunction(MangledName);
+  FunctionType* FT = FunctionType::get(RetTy, ArgTys, false);
+  // ToDo: Some intermediate functions have duplicate names with
+  // different function types. This is OK if the function name
+  // is used internally and finally translated to unique function
+  // names. However it is better to have a way to differentiate
+  // between intermidiate functions and final functions and make
+  // sure final functions have unique names.
+  SPIRVDBG(
+  if (!HasFuncPtrArg && Func && Func->getFunctionType() != FT) {
+    dbgs() << "Warning: Function name conflict:\n"
+       << *Func << '\n'
+       << " => " << *FT << '\n';
+  }
+  )
+  if (!Func || Func->getFunctionType() != FT) {
+    DEBUG(for (auto& I:ArgTys) {
+      dbgs() << *I << '\n';
+    });
+    Func = Function::Create(FT, GlobalValue::ExternalLinkage, MangledName, M);
+    Func->setCallingConv(CallingConv::SPIR_FUNC);
+    if (isFuncNoUnwind())
+      Func->addFnAttr(Attribute::NoUnwind);
+  }
+  auto Call = CallInst::Create(Func,
+      transValue(Ops, BB->getParent(), BB), "", BB);
+  setName(Call, BI);
+  setAttrByCalledFunc(Call);
+  SPIRVDBG(spvdbgs() << "[transInstToBuiltinCall] " << *BI << " -> "; dbgs() <<
+      *Call << '\n';)
+  Instruction *Inst = Call;
+  Inst = transOCLBuiltinPostproc(BI, Call, BB, FuncName);
+  return Inst;
+}
+
+std::string
+SPIRVToLLVM::getOCLBuiltinName(SPIRVInstruction* BI) {
+  auto OC = BI->getOpCode();
+  if (OC == OpGenericCastToPtrExplicit)
+    return getOCLGenericCastToPtrName(BI);
+  if (isCvtOpCode(OC))
+    return getOCLConvertBuiltinName(BI);
+  if (OC == OpBuildNDRange) {
+    auto NDRangeInst = static_cast<SPIRVBuildNDRange *>(BI);
+    auto EleTy = ((NDRangeInst->getOperands())[0])->getType();
+    int Dim = EleTy->isTypeArray() ? EleTy->getArrayLength() : 1;
+    // cygwin does not have std::to_string
+    ostringstream OS;
+    OS << Dim;
+    assert((EleTy->isTypeInt() && Dim == 1) ||
+        (EleTy->isTypeArray() && Dim >= 2 && Dim <= 3));
+    return std::string(kOCLBuiltinName::NDRangePrefix) + OS.str() + "D";
+  }
+  auto Name = OCLSPIRVBuiltinMap::rmap(OC);
+
+  SPIRVType *T = nullptr;
+  switch(OC) {
+  case OpImageRead:
+    T = BI->getType();
+    break;
+  case OpImageWrite:
+    T = BI->getOperands()[2]->getType();
+    break;
+  default:
+    // do nothing
+    break;
+  }
+  if (T && T->isTypeVector())
+    T = T->getVectorComponentType();
+  if (T)
+    Name += T->isTypeFloat()?'f':'i';
+
+  return Name;
+}
+
+Instruction *
+SPIRVToLLVM::transOCLBuiltinFromInst(SPIRVInstruction *BI, BasicBlock *BB) {
+  assert(BB && "Invalid BB");
+  auto FuncName = getOCLBuiltinName(BI);
+  return transBuiltinFromInst(FuncName, BI, BB);
+}
+
+Instruction *
+SPIRVToLLVM::transSPIRVBuiltinFromInst(SPIRVInstruction *BI, BasicBlock *BB) {
+  assert(BB && "Invalid BB");
+  return transBuiltinFromInst(getSPIRVFuncName(BI->getOpCode()), BI, BB);
+}
+
+bool
+SPIRVToLLVM::translate() {
+  if (!transAddressingModel())
+    return false;
+
+  DbgTran.createCompileUnit();
+  DbgTran.addDbgInfoVersion();
+
+  for (unsigned I = 0, E = BM->getNumVariables(); I != E; ++I) {
+    auto BV = BM->getVariable(I);
+    if (BV->getStorageClass() != StorageClassFunction)
+      transValue(BV, nullptr, nullptr);
+  }
+
+  for (unsigned I = 0, E = BM->getNumFunctions(); I != E; ++I) {
+    transFunction(BM->getFunction(I));
+  }
+  if (!transKernelMetadata())
+    return false;
+  if (!transFPContractMetadata())
+    return false;
+  if (!transSourceLanguage())
+    return false;
+  if (!transSourceExtension())
+    return false;
+  transGeneratorMD();
+  if (!transOCLBuiltinsFromVariables())
+    return false;
+  if (!postProcessOCL())
+    return false;
+  eraseUselessFunctions(M);
+  DbgTran.finalize();
+  return true;
+}
+
+bool
+SPIRVToLLVM::transAddressingModel() {
+  switch (BM->getAddressingModel()) {
+  case AddressingModelPhysical64:
+    M->setTargetTriple(SPIR_TARGETTRIPLE64);
+    M->setDataLayout(SPIR_DATALAYOUT64);
+    break;
+  case AddressingModelPhysical32:
+    M->setTargetTriple(SPIR_TARGETTRIPLE32);
+    M->setDataLayout(SPIR_DATALAYOUT32);
+    break;
+  case AddressingModelLogical:
+    // Do not set target triple and data layout
+    break;
+  default:
+    SPIRVCKRT(0, InvalidAddressingModel, "Actual addressing mode is " +
+        (unsigned)BM->getAddressingModel());
+  }
+  return true;
+}
+
+bool
+SPIRVToLLVM::transDecoration(SPIRVValue *BV, Value *V) {
+  if (!transAlign(BV, V))
+    return false;
+  DbgTran.transDbgInfo(BV, V);
+  return true;
+}
+
+bool
+SPIRVToLLVM::transFPContractMetadata() {
+  bool ContractOff = false;
+  for (unsigned I = 0, E = BM->getNumFunctions(); I != E; ++I) {
+    SPIRVFunction *BF = BM->getFunction(I);
+    if (!isOpenCLKernel(BF))
+      continue;
+    if (BF->getExecutionMode(ExecutionModeContractionOff)) {
+      ContractOff = true;
+      break;
+    }
+  }
+  if (!ContractOff)
+    M->getOrInsertNamedMetadata(kSPIR2MD::FPContract);
+  return true;
+}
+
+std::string SPIRVToLLVM::transOCLImageTypeAccessQualifier(
+    SPIRV::SPIRVTypeImage* ST) {
+  return SPIRSPIRVAccessQualifierMap::rmap(ST->getAccessQualifier());
+}
+
+bool
+SPIRVToLLVM::transKernelMetadata() {
+  NamedMDNode *KernelMDs = M->getOrInsertNamedMetadata(SPIR_MD_KERNELS);
+  for (unsigned I = 0, E = BM->getNumFunctions(); I != E; ++I) {
+    SPIRVFunction *BF = BM->getFunction(I);
+    Function *F = static_cast<Function *>(getTranslatedValue(BF));
+    assert(F && "Invalid translated function");
+    if (F->getCallingConv() != CallingConv::SPIR_KERNEL)
+      continue;
+    std::vector<llvm::Metadata*> KernelMD;
+    KernelMD.push_back(ValueAsMetadata::get(F));
+
+    // Generate metadata for kernel_arg_address_spaces
+    addOCLKernelArgumentMetadata(Context, KernelMD,
+        SPIR_MD_KERNEL_ARG_ADDR_SPACE, BF,
+        [=](SPIRVFunctionParameter *Arg){
+      SPIRVType *ArgTy = Arg->getType();
+      SPIRAddressSpace AS = SPIRAS_Private;
+      if (ArgTy->isTypePointer())
+        AS = SPIRSPIRVAddrSpaceMap::rmap(ArgTy->getPointerStorageClass());
+      else if (ArgTy->isTypeOCLImage() || ArgTy->isTypePipe())
+        AS = SPIRAS_Global;
+      return ConstantAsMetadata::get(
+          ConstantInt::get(Type::getInt32Ty(*Context), AS));
+    });
+    // Generate metadata for kernel_arg_access_qual
+    addOCLKernelArgumentMetadata(Context, KernelMD,
+        SPIR_MD_KERNEL_ARG_ACCESS_QUAL, BF,
+        [=](SPIRVFunctionParameter *Arg){
+      std::string Qual;
+      auto T = Arg->getType();
+      if (T->isTypeOCLImage()) {
+        auto ST = static_cast<SPIRVTypeImage *>(T);
+        Qual = transOCLImageTypeAccessQualifier(ST);
+      } else if (T->isTypePipe()){
+        auto PT = static_cast<SPIRVTypePipe *>(T);
+        Qual = transOCLPipeTypeAccessQualifier(PT);
+      } else
+        Qual = "none";
+      return MDString::get(*Context, Qual);
+    });
+    // Generate metadata for kernel_arg_type
+    addOCLKernelArgumentMetadata(Context, KernelMD,
+        SPIR_MD_KERNEL_ARG_TYPE, BF,
+        [=](SPIRVFunctionParameter *Arg){
+      return transOCLKernelArgTypeName(Arg);
+    });
+    // Generate metadata for kernel_arg_type_qual
+    addOCLKernelArgumentMetadata(Context, KernelMD,
+        SPIR_MD_KERNEL_ARG_TYPE_QUAL, BF,
+        [=](SPIRVFunctionParameter *Arg){
+      std::string Qual;
+      if (Arg->hasDecorate(DecorationVolatile))
+        Qual = kOCLTypeQualifierName::Volatile;
+      Arg->foreachAttr([&](SPIRVFuncParamAttrKind Kind){
+        Qual += Qual.empty() ? "" : " ";
+        switch(Kind){
+        case FunctionParameterAttributeNoAlias:
+          Qual += kOCLTypeQualifierName::Restrict;
+          break;
+        case FunctionParameterAttributeNoWrite:
+          Qual += kOCLTypeQualifierName::Const;
+          break;
+        default:
+          // do nothing.
+          break;
+        }
+      });
+      if (Arg->getType()->isTypePipe()) {
+        Qual += Qual.empty() ? "" : " ";
+        Qual += kOCLTypeQualifierName::Pipe;
+      }
+      return MDString::get(*Context, Qual);
+    });
+    // Generate metadata for kernel_arg_base_type
+    addOCLKernelArgumentMetadata(Context, KernelMD,
+        SPIR_MD_KERNEL_ARG_BASE_TYPE, BF,
+        [=](SPIRVFunctionParameter *Arg){
+      return transOCLKernelArgTypeName(Arg);
+    });
+    // Generate metadata for kernel_arg_name
+    if (SPIRVGenKernelArgNameMD) {
+      bool ArgHasName = true;
+      BF->foreachArgument([&](SPIRVFunctionParameter *Arg){
+        ArgHasName &= !Arg->getName().empty();
+      });
+      if (ArgHasName)
+        addOCLKernelArgumentMetadata(Context, KernelMD,
+            SPIR_MD_KERNEL_ARG_NAME, BF,
+            [=](SPIRVFunctionParameter *Arg){
+          return MDString::get(*Context, Arg->getName());
+        });
+    }
+    // Generate metadata for reqd_work_group_size
+    if (auto EM = BF->getExecutionMode(ExecutionModeLocalSize)) {
+      KernelMD.push_back(getMDNodeStringIntVec(Context,
+          kSPIR2MD::WGSize, EM->getLiterals()));
+    }
+    // Generate metadata for work_group_size_hint
+    if (auto EM = BF->getExecutionMode(ExecutionModeLocalSizeHint)) {
+      KernelMD.push_back(getMDNodeStringIntVec(Context,
+          kSPIR2MD::WGSizeHint, EM->getLiterals()));
+    }
+    // Generate metadata for vec_type_hint
+    if (auto EM = BF->getExecutionMode(ExecutionModeVecTypeHint)) {
+      std::vector<Metadata*> MetadataVec;
+      MetadataVec.push_back(MDString::get(*Context, kSPIR2MD::VecTyHint));
+      Type *VecHintTy = decodeVecTypeHint(*Context, EM->getLiterals()[0]);
       assert(VecHintTy);
-      MetadataVec.push_back(ValueAsMetadata::get(UndefValue::get(VecHintTy)));
-      MetadataVec.push_back(
-          ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(*Context),
-              1)));
-      KernelMD.push_back(MDNode::get(*Context, MetadataVec));
-    }
-
-    llvm::MDNode *Node = MDNode::get(*Context, KernelMD);
-    KernelMDs->addOperand(Node);
-  }
-  return true;
-}
-
-bool
-SPIRVToLLVM::transAlign(SPIRVValue *BV, Value *V) {
-  if (auto AL = dyn_cast<AllocaInst>(V)) {
-    SPIRVWord Align = 0;
-    if (BV->hasAlignment(&Align))
-      AL->setAlignment(Align);
-    return true;
-  }
-  if (auto GV = dyn_cast<GlobalVariable>(V)) {
-    SPIRVWord Align = 0;
-    if (BV->hasAlignment(&Align))
-      GV->setAlignment(Align);
-    return true;
-  }
-  return true;
-}
-
-void
-SPIRVToLLVM::transOCLVectorLoadStore(std::string& UnmangledName,
-    std::vector<SPIRVWord> &BArgs) {
-  if (UnmangledName.find("vload") == 0 &&
-      UnmangledName.find("n") != std::string::npos) {
-    if (BArgs.back() != 1) {
-      std::stringstream SS;
-      SS << BArgs.back();
-      UnmangledName.replace(UnmangledName.find("n"), 1, SS.str());
-    } else {
-      UnmangledName.erase(UnmangledName.find("n"), 1);
-    }
-    BArgs.pop_back();
-  } else if (UnmangledName.find("vstore") == 0) {
-    if (UnmangledName.find("n") != std::string::npos) {
-      auto T = BM->getValueType(BArgs[0]);
-      if (T->isTypeVector()) {
-        auto W = T->getVectorComponentCount();
-        std::stringstream SS;
-        SS << W;
-        UnmangledName.replace(UnmangledName.find("n"), 1, SS.str());
-      } else {
-        UnmangledName.erase(UnmangledName.find("n"), 1);
-      }
-    }
-    if (UnmangledName.find("_r") != std::string::npos) {
-      UnmangledName.replace(UnmangledName.find("_r"), 2, std::string("_") +
-          SPIRSPIRVFPRoundingModeMap::rmap(static_cast<SPIRVFPRoundingModeKind>(
-              BArgs.back())));
-      BArgs.pop_back();
-    }
-   }
-}
-
-// printf is not mangled. The function type should have just one argument.
-// read_image*: the second argument should be mangled as sampler.
-Instruction *
-SPIRVToLLVM::transOCLBuiltinFromExtInst(SPIRVExtInst *BC, BasicBlock *BB) {
-  assert(BB && "Invalid BB");
-  std::string MangledName;
-  SPIRVWord EntryPoint = BC->getExtOp();
-  SPIRVExtInstSetKind Set = BM->getBuiltinSet(BC->getExtSetId());
-  bool IsVarArg = false;
-  bool IsPrintf = false;
-  std::string UnmangledName;
-  auto BArgs = BC->getArguments();
-
-  assert (Set == SPIRVEIS_OpenCL && "Not OpenCL extended instruction");
-  if (EntryPoint == OpenCLLIB::Printf)
-    IsPrintf = true;
-  else {
-    UnmangledName = OCLExtOpMap::map(static_cast<OCLExtOpKind>(
-        EntryPoint));
-  }
-
-  SPIRVDBG(spvdbgs() << "[transOCLBuiltinFromExtInst] OrigUnmangledName: " <<
-      UnmangledName << '\n');
-  transOCLVectorLoadStore(UnmangledName, BArgs);
-
-  std::vector<Type *> ArgTypes = transTypeVector(BC->getValueTypes(BArgs));
-
-  if (IsPrintf) {
-    MangledName = "printf";
-    IsVarArg = true;
-    ArgTypes.resize(1);
-  } else if (UnmangledName.find("read_image") == 0) {
-    auto ModifiedArgTypes = ArgTypes;
-    ModifiedArgTypes[1] = getOrCreateOpaquePtrType(M, "opencl.sampler_t");
-    MangleOpenCLBuiltin(UnmangledName, ModifiedArgTypes, MangledName);
-  } else {
-    MangleOpenCLBuiltin(UnmangledName, ArgTypes, MangledName);
-  }
-  SPIRVDBG(spvdbgs() << "[transOCLBuiltinFromExtInst] ModifiedUnmangledName: " <<
-      UnmangledName << " MangledName: " << MangledName << '\n');
-
-  FunctionType *FT = FunctionType::get(
-      transType(BC->getType()),
-      ArgTypes,
-      IsVarArg);
-  Function *F = M->getFunction(MangledName);
-  if (!F) {
-    F = Function::Create(FT,
-      GlobalValue::ExternalLinkage,
-      MangledName,
-      M);
-    F->setCallingConv(CallingConv::SPIR_FUNC);
-    if (isFuncNoUnwind())
-      F->addFnAttr(Attribute::NoUnwind);
-  }
-  auto Args = transValue(BC->getValues(BArgs), F, BB);
-  SPIRVDBG(dbgs() << "[transOCLBuiltinFromExtInst] Function: " << *F <<
-      ", Args: ";
-    for (auto &I:Args) dbgs() << *I << ", "; dbgs() << '\n');
-  CallInst *Call = CallInst::Create(F,
-      Args,
-      BC->getName(),
-      BB);
-  setCallingConv(Call);
-  addFnAttr(Context, Call, Attribute::NoUnwind);
-  return transOCLBuiltinPostproc(BC, Call, BB, UnmangledName);
-}
-
-Instruction *
-SPIRVToLLVM::transOCLBarrierFence(SPIRVInstruction* MB, BasicBlock *BB) {
-  assert(BB && "Invalid BB");
-  std::string FuncName;
-  auto getIntVal = [](SPIRVValue *value){
-    return static_cast<SPIRVConstant*>(value)->getZExtIntValue();
-  };
-  SPIRVWord MemSema = 0;
-  if (MB->getOpCode() == OpMemoryBarrier) {
-    auto MemB = static_cast<SPIRVMemoryBarrier*>(MB);
-    FuncName = "mem_fence";
-    MemSema = getIntVal(MemB->getOpValue(1));
-  } else if (MB->getOpCode() == OpControlBarrier) {
-    auto CtlB = static_cast<SPIRVControlBarrier*>(MB);
-    SPIRVWord Ver = 1;
-    BM->getSourceLanguage(&Ver);
-    FuncName = (Ver <= 12) ? kOCLBuiltinName::Barrier :
-        kOCLBuiltinName::WorkGroupBarrier;
-    MemSema = getIntVal(CtlB->getMemSemantic());
-  } else {
-    llvm_unreachable("Invalid instruction");
-  }
-  std::string MangledName;
-  Type* Int32Ty = Type::getInt32Ty(*Context);
-  Type* VoidTy = Type::getVoidTy(*Context);
-  Type* ArgTy[] = {Int32Ty};
-  MangleOpenCLBuiltin(FuncName, ArgTy, MangledName);
-  Function *Func = M->getFunction(MangledName);
-  if (!Func) {
-    FunctionType *FT = FunctionType::get(VoidTy, ArgTy, false);
-    Func = Function::Create(FT, GlobalValue::ExternalLinkage, MangledName, M);
-    Func->setCallingConv(CallingConv::SPIR_FUNC);
-    if (isFuncNoUnwind())
-      Func->addFnAttr(Attribute::NoUnwind);
-  }
-  Value *Arg[] = {ConstantInt::get(Int32Ty,
-      rmapBitMask<OCLMemFenceMap>(MemSema))};
-  auto Call = CallInst::Create(Func, Arg, "", BB);
-  setName(Call, MB);
-  setAttrByCalledFunc(Call);
-  SPIRVDBG(spvdbgs() << "[transBarrier] " << *MB << " -> ";
-    dbgs() << *Call << '\n';)
-  return Call;
-}
-
-// SPIR-V only contains language version. Use OpenCL language version as
-// SPIR version.
-bool
-SPIRVToLLVM::transSourceLanguage() {
-  SPIRVWord Ver = 0;
-  SourceLanguage Lang = BM->getSourceLanguage(&Ver);
-  assert((Lang == SourceLanguageOpenCL_C ||
-      Lang == SourceLanguageOpenCL_CPP) && "Unsupported source language");
-  unsigned short Major = 0;
-  unsigned char Minor = 0;
-  unsigned char Rev = 0;
-  std::tie(Major, Minor, Rev) = decodeOCLVer(Ver);
-  SPIRVMDBuilder Builder(*M);
-  Builder.addNamedMD(kSPIRVMD::Source)
-            .addOp()
-              .add(Lang)
-              .add(Ver)
-              .done();
-  // ToDo: Phasing out usage of old SPIR metadata
-  if (Ver <= kOCLVer::CL12)
-    addOCLVersionMetadata(Context, M, kSPIR2MD::SPIRVer, 1, 2);
-  else
-    addOCLVersionMetadata(Context, M, kSPIR2MD::SPIRVer, 2, 0);
-
-  addOCLVersionMetadata(Context, M, kSPIR2MD::OCLVer, Major, Minor);
-  return true;
-}
-
-bool
-SPIRVToLLVM::transSourceExtension() {
-  auto ExtSet = rmap<OclExt::Kind>(BM->getExtension());
-  auto CapSet = rmap<OclExt::Kind>(BM->getCapability());
-  ExtSet.insert(CapSet.begin(), CapSet.end());
-  auto OCLExtensions = map<std::string>(ExtSet);
-  std::set<std::string> OCLOptionalCoreFeatures;
-  static const char *OCLOptCoreFeatureNames[] = {
-      "cl_images", "cl_doubles",
-  };
-  for (auto &I : OCLOptCoreFeatureNames) {
-    auto Loc = OCLExtensions.find(I);
-    if (Loc != OCLExtensions.end()) {
-      OCLExtensions.erase(Loc);
-      OCLOptionalCoreFeatures.insert(I);
-    }
-  }
-  addNamedMetadataStringSet(Context, M, kSPIR2MD::Extensions, OCLExtensions);
-  addNamedMetadataStringSet(Context, M, kSPIR2MD::OptFeatures,
-                            OCLOptionalCoreFeatures);
-  return true;
-}
-
-// If the argument is unsigned return uconvert*, otherwise return convert*.
-std::string
-SPIRVToLLVM::getOCLConvertBuiltinName(SPIRVInstruction* BI) {
-  auto OC = BI->getOpCode();
-  assert(isCvtOpCode(OC) && "Not convert instruction");
-  auto U = static_cast<SPIRVUnary *>(BI);
-  std::string Name;
-  if (isCvtFromUnsignedOpCode(OC))
-    Name = "u";
-  Name += "convert_";
-  Name += mapSPIRVTypeToOCLType(U->getType(),
-      !isCvtToUnsignedOpCode(OC));
-  SPIRVFPRoundingModeKind Rounding;
-  if (U->isSaturatedConversion())
-    Name += "_sat";
-  if (U->hasFPRoundingMode(&Rounding)) {
-    Name += "_";
-    Name += SPIRSPIRVFPRoundingModeMap::rmap(Rounding);
-  }
-  return Name;
-}
-
-//Check Address Space of the Pointer Type
-std::string
-SPIRVToLLVM::getOCLGenericCastToPtrName(SPIRVInstruction* BI) {
-  auto GenericCastToPtrInst = BI->getType()->getPointerStorageClass();
-  switch (GenericCastToPtrInst) {
-    case StorageClassCrossWorkgroup:
-      return std::string(kOCLBuiltinName::ToGlobal);
-    case StorageClassWorkgroup:
-      return std::string(kOCLBuiltinName::ToLocal);
-    case StorageClassFunction:
-      return std::string(kOCLBuiltinName::ToPrivate);
-    default:
-      llvm_unreachable("Invalid address space");
-      return "";
-  }
-}
-
-llvm::GlobalValue::LinkageTypes
-SPIRVToLLVM::transLinkageType(const SPIRVValue* V) {
-  if (V->getLinkageType() == LinkageTypeInternal) {
-    return GlobalValue::InternalLinkage;
-  }
-  else if (V->getLinkageType() == LinkageTypeImport) {
-    // Function declaration
-    if (V->getOpCode() == OpFunction) {
-      if (static_cast<const SPIRVFunction*>(V)->getNumBasicBlock() == 0)
-        return GlobalValue::ExternalLinkage;
-    }
-    // Variable declaration
-    if (V->getOpCode() == OpVariable) {
-      if (static_cast<const SPIRVVariable*>(V)->getInitializer() == 0)
-        return GlobalValue::ExternalLinkage;
-    }
-    // Definition
-    return GlobalValue::AvailableExternallyLinkage;
-  }
-  else {// LinkageTypeExport
-    if (V->getOpCode() == OpVariable) {
-      if (static_cast<const SPIRVVariable*>(V)->getInitializer() == 0 )
-        // Tentative definition
-        return GlobalValue::CommonLinkage;
-    }
-    return GlobalValue::ExternalLinkage;
-  }
-}
-
-Instruction *SPIRVToLLVM::transOCLAllAny(SPIRVInstruction *I, BasicBlock *BB) {
-  CallInst *CI = cast<CallInst>(transSPIRVBuiltinFromInst(I, BB));
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  return cast<Instruction>(mapValue(
-      I, mutateCallInstOCL(
-             M, CI,
-             [=](CallInst *, std::vector<Value *> &Args, llvm::Type *&RetTy) {
-               Type *Int32Ty = Type::getInt32Ty(*Context);
-               auto OldArg = CI->getOperand(0);
-               auto NewArgTy = VectorType::get(
-                   Int32Ty, OldArg->getType()->getVectorNumElements());
-               auto NewArg =
-                   CastInst::CreateSExtOrBitCast(OldArg, NewArgTy, "", CI);
-               Args[0] = NewArg;
-               RetTy = Int32Ty;
-               return CI->getCalledFunction()->getName();
-             },
-             [=](CallInst *NewCI) -> Instruction * {
-               return CastInst::CreateTruncOrBitCast(
-                   NewCI, Type::getInt1Ty(*Context), "", NewCI->getNextNode());
-             },
-             &Attrs)));
-}
-
-Instruction *SPIRVToLLVM::transOCLRelational(SPIRVInstruction *I, BasicBlock *BB) {
-  CallInst *CI = cast<CallInst>(transSPIRVBuiltinFromInst(I, BB));
-  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
-  return cast<Instruction>(mapValue(
-      I, mutateCallInstOCL(
-             M, CI,
-             [=](CallInst *, std::vector<Value *> &Args, llvm::Type *&RetTy) {
-               Type *IntTy = Type::getInt32Ty(*Context);
-               RetTy = IntTy;
-               if (CI->getType()->isVectorTy()) {
-                 if (cast<VectorType>(CI->getOperand(0)->getType())
-                         ->getElementType()
-                         ->isDoubleTy())
-                   IntTy = Type::getInt64Ty(*Context);
-                 if (cast<VectorType>(CI->getOperand(0)->getType())
-                         ->getElementType()
-                         ->isHalfTy())
-                   IntTy = Type::getInt16Ty(*Context);
-                 RetTy = VectorType::get(IntTy,
-                                         CI->getType()->getVectorNumElements());
-               }
-               return CI->getCalledFunction()->getName();
-             },
-             [=](CallInst *NewCI) -> Instruction * {
-               Type *RetTy = Type::getInt1Ty(*Context);
-               if (NewCI->getType()->isVectorTy())
-                 RetTy =
-                     VectorType::get(Type::getInt1Ty(*Context),
-                                     NewCI->getType()->getVectorNumElements());
-               return CastInst::CreateTruncOrBitCast(NewCI, RetTy, "",
-                                                     NewCI->getNextNode());
-             },
-             &Attrs)));
-}
-}
-
-bool
-llvm::ReadSPIRV(LLVMContext &C, std::istream &IS, Module *&M,
-    std::string &ErrMsg) {
-  M = new Module("", C);
-  std::unique_ptr<SPIRVModule> BM(SPIRVModule::createSPIRVModule());
-
-  BM->setAutoAddCapability(false);
-  IS >> *BM;
-
-  SPIRVToLLVM BTL(M, BM.get());
-  bool Succeed = true;
-  if (!BTL.translate()) {
-    BM->getError(ErrMsg);
-    Succeed = false;
-  }
-  PassManager PassMgr;
-  PassMgr.add(createSPIRVToOCL20());
-  PassMgr.add(createOCL20To12());
-  PassMgr.run(*M);
-
-  if (DbgSaveTmpLLVM)
-    dumpLLVM(M, DbgTmpLLVMFileName);
-  if (!Succeed) {
-    delete M;
-    M = nullptr;
-  }
-  return Succeed;
-}
+      MetadataVec.push_back(ValueAsMetadata::get(UndefValue::get(VecHintTy)));
+      MetadataVec.push_back(
+          ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(*Context),
+              1)));
+      KernelMD.push_back(MDNode::get(*Context, MetadataVec));
+    }
+
+    llvm::MDNode *Node = MDNode::get(*Context, KernelMD);
+    KernelMDs->addOperand(Node);
+  }
+  return true;
+}
+
+bool
+SPIRVToLLVM::transAlign(SPIRVValue *BV, Value *V) {
+  if (auto AL = dyn_cast<AllocaInst>(V)) {
+    SPIRVWord Align = 0;
+    if (BV->hasAlignment(&Align))
+      AL->setAlignment(Align);
+    return true;
+  }
+  if (auto GV = dyn_cast<GlobalVariable>(V)) {
+    SPIRVWord Align = 0;
+    if (BV->hasAlignment(&Align))
+      GV->setAlignment(Align);
+    return true;
+  }
+  return true;
+}
+
+void
+SPIRVToLLVM::transOCLVectorLoadStore(std::string& UnmangledName,
+    std::vector<SPIRVWord> &BArgs) {
+  if (UnmangledName.find("vload") == 0 &&
+      UnmangledName.find("n") != std::string::npos) {
+    if (BArgs.back() != 1) {
+      std::stringstream SS;
+      SS << BArgs.back();
+      UnmangledName.replace(UnmangledName.find("n"), 1, SS.str());
+    } else {
+      UnmangledName.erase(UnmangledName.find("n"), 1);
+    }
+    BArgs.pop_back();
+  } else if (UnmangledName.find("vstore") == 0) {
+    if (UnmangledName.find("n") != std::string::npos) {
+      auto T = BM->getValueType(BArgs[0]);
+      if (T->isTypeVector()) {
+        auto W = T->getVectorComponentCount();
+        std::stringstream SS;
+        SS << W;
+        UnmangledName.replace(UnmangledName.find("n"), 1, SS.str());
+      } else {
+        UnmangledName.erase(UnmangledName.find("n"), 1);
+      }
+    }
+    if (UnmangledName.find("_r") != std::string::npos) {
+      UnmangledName.replace(UnmangledName.find("_r"), 2, std::string("_") +
+          SPIRSPIRVFPRoundingModeMap::rmap(static_cast<SPIRVFPRoundingModeKind>(
+              BArgs.back())));
+      BArgs.pop_back();
+    }
+   }
+}
+
+// printf is not mangled. The function type should have just one argument.
+// read_image*: the second argument should be mangled as sampler.
+Instruction *
+SPIRVToLLVM::transOCLBuiltinFromExtInst(SPIRVExtInst *BC, BasicBlock *BB) {
+  assert(BB && "Invalid BB");
+  std::string MangledName;
+  SPIRVWord EntryPoint = BC->getExtOp();
+  SPIRVExtInstSetKind Set = BM->getBuiltinSet(BC->getExtSetId());
+  bool IsVarArg = false;
+  bool IsPrintf = false;
+  std::string UnmangledName;
+  auto BArgs = BC->getArguments();
+
+  assert (Set == SPIRVEIS_OpenCL && "Not OpenCL extended instruction");
+  if (EntryPoint == OpenCLLIB::Printf)
+    IsPrintf = true;
+  else {
+    UnmangledName = OCLExtOpMap::map(static_cast<OCLExtOpKind>(
+        EntryPoint));
+  }
+
+  SPIRVDBG(spvdbgs() << "[transOCLBuiltinFromExtInst] OrigUnmangledName: " <<
+      UnmangledName << '\n');
+  transOCLVectorLoadStore(UnmangledName, BArgs);
+
+  std::vector<Type *> ArgTypes = transTypeVector(BC->getValueTypes(BArgs));
+
+  if (IsPrintf) {
+    MangledName = "printf";
+    IsVarArg = true;
+    ArgTypes.resize(1);
+  } else if (UnmangledName.find("read_image") == 0) {
+    auto ModifiedArgTypes = ArgTypes;
+    ModifiedArgTypes[1] = getOrCreateOpaquePtrType(M, "opencl.sampler_t");
+    MangleOpenCLBuiltin(UnmangledName, ModifiedArgTypes, MangledName);
+  } else {
+    MangleOpenCLBuiltin(UnmangledName, ArgTypes, MangledName);
+  }
+  SPIRVDBG(spvdbgs() << "[transOCLBuiltinFromExtInst] ModifiedUnmangledName: " <<
+      UnmangledName << " MangledName: " << MangledName << '\n');
+
+  FunctionType *FT = FunctionType::get(
+      transType(BC->getType()),
+      ArgTypes,
+      IsVarArg);
+  Function *F = M->getFunction(MangledName);
+  if (!F) {
+    F = Function::Create(FT,
+      GlobalValue::ExternalLinkage,
+      MangledName,
+      M);
+    F->setCallingConv(CallingConv::SPIR_FUNC);
+    if (isFuncNoUnwind())
+      F->addFnAttr(Attribute::NoUnwind);
+  }
+  auto Args = transValue(BC->getValues(BArgs), F, BB);
+  SPIRVDBG(dbgs() << "[transOCLBuiltinFromExtInst] Function: " << *F <<
+      ", Args: ";
+    for (auto &I:Args) dbgs() << *I << ", "; dbgs() << '\n');
+  CallInst *Call = CallInst::Create(F,
+      Args,
+      BC->getName(),
+      BB);
+  setCallingConv(Call);
+  addFnAttr(Context, Call, Attribute::NoUnwind);
+  return transOCLBuiltinPostproc(BC, Call, BB, UnmangledName);
+}
+
+Instruction *
+SPIRVToLLVM::transOCLBarrierFence(SPIRVInstruction* MB, BasicBlock *BB) {
+  assert(BB && "Invalid BB");
+  std::string FuncName;
+  auto getIntVal = [](SPIRVValue *value){
+    return static_cast<SPIRVConstant*>(value)->getZExtIntValue();
+  };
+  SPIRVWord MemSema = 0;
+  if (MB->getOpCode() == OpMemoryBarrier) {
+    auto MemB = static_cast<SPIRVMemoryBarrier*>(MB);
+    FuncName = "mem_fence";
+    MemSema = getIntVal(MemB->getOpValue(1));
+  } else if (MB->getOpCode() == OpControlBarrier) {
+    auto CtlB = static_cast<SPIRVControlBarrier*>(MB);
+    SPIRVWord Ver = 1;
+    BM->getSourceLanguage(&Ver);
+    FuncName = (Ver <= 12) ? kOCLBuiltinName::Barrier :
+        kOCLBuiltinName::WorkGroupBarrier;
+    MemSema = getIntVal(CtlB->getMemSemantic());
+  } else {
+    llvm_unreachable("Invalid instruction");
+  }
+  std::string MangledName;
+  Type* Int32Ty = Type::getInt32Ty(*Context);
+  Type* VoidTy = Type::getVoidTy(*Context);
+  Type* ArgTy[] = {Int32Ty};
+  MangleOpenCLBuiltin(FuncName, ArgTy, MangledName);
+  Function *Func = M->getFunction(MangledName);
+  if (!Func) {
+    FunctionType *FT = FunctionType::get(VoidTy, ArgTy, false);
+    Func = Function::Create(FT, GlobalValue::ExternalLinkage, MangledName, M);
+    Func->setCallingConv(CallingConv::SPIR_FUNC);
+    if (isFuncNoUnwind())
+      Func->addFnAttr(Attribute::NoUnwind);
+  }
+  Value *Arg[] = {ConstantInt::get(Int32Ty,
+      rmapBitMask<OCLMemFenceMap>(MemSema))};
+  auto Call = CallInst::Create(Func, Arg, "", BB);
+  setName(Call, MB);
+  setAttrByCalledFunc(Call);
+  SPIRVDBG(spvdbgs() << "[transBarrier] " << *MB << " -> ";
+    dbgs() << *Call << '\n';)
+  return Call;
+}
+
+// SPIR-V only contains language version. Use OpenCL language version as
+// SPIR version.
+bool
+SPIRVToLLVM::transSourceLanguage() {
+  SPIRVWord Ver = 0;
+  SourceLanguage Lang = BM->getSourceLanguage(&Ver);
+  assert((Lang == SourceLanguageOpenCL_C ||
+      Lang == SourceLanguageOpenCL_CPP) && "Unsupported source language");
+  unsigned short Major = 0;
+  unsigned char Minor = 0;
+  unsigned char Rev = 0;
+  std::tie(Major, Minor, Rev) = decodeOCLVer(Ver);
+  SPIRVMDBuilder Builder(*M);
+  Builder.addNamedMD(kSPIRVMD::Source)
+            .addOp()
+              .add(Lang)
+              .add(Ver)
+              .done();
+  // ToDo: Phasing out usage of old SPIR metadata
+  if (Ver <= kOCLVer::CL12)
+    addOCLVersionMetadata(Context, M, kSPIR2MD::SPIRVer, 1, 2);
+  else
+    addOCLVersionMetadata(Context, M, kSPIR2MD::SPIRVer, 2, 0);
+
+  addOCLVersionMetadata(Context, M, kSPIR2MD::OCLVer, Major, Minor);
+  return true;
+}
+
+bool
+SPIRVToLLVM::transSourceExtension() {
+  auto ExtSet = rmap<OclExt::Kind>(BM->getExtension());
+  auto CapSet = rmap<OclExt::Kind>(BM->getCapability());
+  ExtSet.insert(CapSet.begin(), CapSet.end());
+  auto OCLExtensions = map<std::string>(ExtSet);
+  std::set<std::string> OCLOptionalCoreFeatures;
+  static const char *OCLOptCoreFeatureNames[] = {
+      "cl_images", "cl_doubles",
+  };
+  for (auto &I : OCLOptCoreFeatureNames) {
+    auto Loc = OCLExtensions.find(I);
+    if (Loc != OCLExtensions.end()) {
+      OCLExtensions.erase(Loc);
+      OCLOptionalCoreFeatures.insert(I);
+    }
+  }
+  addNamedMetadataStringSet(Context, M, kSPIR2MD::Extensions, OCLExtensions);
+  addNamedMetadataStringSet(Context, M, kSPIR2MD::OptFeatures,
+                            OCLOptionalCoreFeatures);
+  return true;
+}
+
+// If the argument is unsigned return uconvert*, otherwise return convert*.
+std::string
+SPIRVToLLVM::getOCLConvertBuiltinName(SPIRVInstruction* BI) {
+  auto OC = BI->getOpCode();
+  assert(isCvtOpCode(OC) && "Not convert instruction");
+  auto U = static_cast<SPIRVUnary *>(BI);
+  std::string Name;
+  if (isCvtFromUnsignedOpCode(OC))
+    Name = "u";
+  Name += "convert_";
+  Name += mapSPIRVTypeToOCLType(U->getType(),
+      !isCvtToUnsignedOpCode(OC));
+  SPIRVFPRoundingModeKind Rounding;
+  if (U->isSaturatedConversion())
+    Name += "_sat";
+  if (U->hasFPRoundingMode(&Rounding)) {
+    Name += "_";
+    Name += SPIRSPIRVFPRoundingModeMap::rmap(Rounding);
+  }
+  return Name;
+}
+
+//Check Address Space of the Pointer Type
+std::string
+SPIRVToLLVM::getOCLGenericCastToPtrName(SPIRVInstruction* BI) {
+  auto GenericCastToPtrInst = BI->getType()->getPointerStorageClass();
+  switch (GenericCastToPtrInst) {
+    case StorageClassCrossWorkgroup:
+      return std::string(kOCLBuiltinName::ToGlobal);
+    case StorageClassWorkgroup:
+      return std::string(kOCLBuiltinName::ToLocal);
+    case StorageClassFunction:
+      return std::string(kOCLBuiltinName::ToPrivate);
+    default:
+      llvm_unreachable("Invalid address space");
+      return "";
+  }
+}
+
+llvm::GlobalValue::LinkageTypes
+SPIRVToLLVM::transLinkageType(const SPIRVValue* V) {
+  if (V->getLinkageType() == LinkageTypeInternal) {
+    return GlobalValue::InternalLinkage;
+  }
+  else if (V->getLinkageType() == LinkageTypeImport) {
+    // Function declaration
+    if (V->getOpCode() == OpFunction) {
+      if (static_cast<const SPIRVFunction*>(V)->getNumBasicBlock() == 0)
+        return GlobalValue::ExternalLinkage;
+    }
+    // Variable declaration
+    if (V->getOpCode() == OpVariable) {
+      if (static_cast<const SPIRVVariable*>(V)->getInitializer() == 0)
+        return GlobalValue::ExternalLinkage;
+    }
+    // Definition
+    return GlobalValue::AvailableExternallyLinkage;
+  }
+  else {// LinkageTypeExport
+    if (V->getOpCode() == OpVariable) {
+      if (static_cast<const SPIRVVariable*>(V)->getInitializer() == 0 )
+        // Tentative definition
+        return GlobalValue::CommonLinkage;
+    }
+    return GlobalValue::ExternalLinkage;
+  }
+}
+
+Instruction *SPIRVToLLVM::transOCLAllAny(SPIRVInstruction *I, BasicBlock *BB) {
+  CallInst *CI = cast<CallInst>(transSPIRVBuiltinFromInst(I, BB));
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  return cast<Instruction>(mapValue(
+      I, mutateCallInstOCL(
+             M, CI,
+             [=](CallInst *, std::vector<Value *> &Args, llvm::Type *&RetTy) {
+               Type *Int32Ty = Type::getInt32Ty(*Context);
+               auto OldArg = CI->getOperand(0);
+               auto NewArgTy = VectorType::get(
+                   Int32Ty, OldArg->getType()->getVectorNumElements());
+               auto NewArg =
+                   CastInst::CreateSExtOrBitCast(OldArg, NewArgTy, "", CI);
+               Args[0] = NewArg;
+               RetTy = Int32Ty;
+               return CI->getCalledFunction()->getName();
+             },
+             [=](CallInst *NewCI) -> Instruction * {
+               return CastInst::CreateTruncOrBitCast(
+                   NewCI, Type::getInt1Ty(*Context), "", NewCI->getNextNode());
+             },
+             &Attrs)));
+}
+
+Instruction *SPIRVToLLVM::transOCLRelational(SPIRVInstruction *I, BasicBlock *BB) {
+  CallInst *CI = cast<CallInst>(transSPIRVBuiltinFromInst(I, BB));
+  AttributeSet Attrs = CI->getCalledFunction()->getAttributes();
+  return cast<Instruction>(mapValue(
+      I, mutateCallInstOCL(
+             M, CI,
+             [=](CallInst *, std::vector<Value *> &Args, llvm::Type *&RetTy) {
+               Type *IntTy = Type::getInt32Ty(*Context);
+               RetTy = IntTy;
+               if (CI->getType()->isVectorTy()) {
+                 if (cast<VectorType>(CI->getOperand(0)->getType())
+                         ->getElementType()
+                         ->isDoubleTy())
+                   IntTy = Type::getInt64Ty(*Context);
+                 if (cast<VectorType>(CI->getOperand(0)->getType())
+                         ->getElementType()
+                         ->isHalfTy())
+                   IntTy = Type::getInt16Ty(*Context);
+                 RetTy = VectorType::get(IntTy,
+                                         CI->getType()->getVectorNumElements());
+               }
+               return CI->getCalledFunction()->getName();
+             },
+             [=](CallInst *NewCI) -> Instruction * {
+               Type *RetTy = Type::getInt1Ty(*Context);
+               if (NewCI->getType()->isVectorTy())
+                 RetTy =
+                     VectorType::get(Type::getInt1Ty(*Context),
+                                     NewCI->getType()->getVectorNumElements());
+               return CastInst::CreateTruncOrBitCast(NewCI, RetTy, "",
+                                                     NewCI->getNextNode());
+             },
+             &Attrs)));
+}
+}
+
+bool
+llvm::ReadSPIRV(LLVMContext &C, std::istream &IS, Module *&M,
+    std::string &ErrMsg) {
+  M = new Module("", C);
+  std::unique_ptr<SPIRVModule> BM(SPIRVModule::createSPIRVModule());
+
+  BM->setAutoAddCapability(false);
+  IS >> *BM;
+
+  SPIRVToLLVM BTL(M, BM.get());
+  bool Succeed = true;
+  if (!BTL.translate()) {
+    BM->getError(ErrMsg);
+    Succeed = false;
+  }
+  legacy::PassManager PassMgr;
+  PassMgr.add(createSPIRVToOCL20());
+  PassMgr.add(createOCL20To12());
+  PassMgr.run(*M);
+
+  if (DbgSaveTmpLLVM)
+    dumpLLVM(M, DbgTmpLLVMFileName);
+  if (!Succeed) {
+    delete M;
+    M = nullptr;
+  }
+  return Succeed;
+}
diff --git a/SPIRVRegularizeLLVM.cpp b/SPIRVRegularizeLLVM.cpp
index 1631b61..b12af32 100644
--- a/SPIRVRegularizeLLVM.cpp
+++ b/SPIRVRegularizeLLVM.cpp
@@ -116,7 +116,7 @@ SPIRVRegularizeLLVM::regularize() {
   //lowerConstantExpressions();
 
   for (auto I = M->begin(), E = M->end(); I != E;) {
-    Function *F = I++;
+    Function *F = &*I++;
     if (F->isDeclaration() && F->use_empty()) {
       F->eraseFromParent();
       continue;
@@ -193,7 +193,7 @@ void
 SPIRVRegularizeLLVM::lowerFuncPtr(Module* M) {
   std::vector<std::pair<Function *, Op>> Work;
   for (auto I = M->begin(), E = M->end(); I != E;) {
-    Function *F = I++;
+    Function *F = &*I++;
     auto AI = F->arg_begin();
     if (hasFunctionPointerArg(F, AI)) {
       auto OC = getSPIRVFuncOC(F->getName());
diff --git a/SPIRVToOCL20.cpp b/SPIRVToOCL20.cpp
index d994490..528cfe9 100644
--- a/SPIRVToOCL20.cpp
+++ b/SPIRVToOCL20.cpp
@@ -308,7 +308,7 @@ void SPIRVToOCL20::visitCallSPRIVImageQuerySize(CallInst *CI) {
   if (imgArray) {
     assert((imgDim == 1 || imgDim == 2) && "invalid image array type");
     // Insert get_image_array_size to the last position of the resulting vector.
-    Type * sizeTy = Type::getIntNTy(*Ctx, M->getDataLayout()->getPointerSizeInBits(0));
+    Type * sizeTy = Type::getIntNTy(*Ctx, M->getDataLayout().getPointerSizeInBits(0));
     Instruction * getImageArraySize =
       addCallInst(M, kOCLBuiltinName::GetImageArraySize, sizeTy,
                   CI->getArgOperand(0), &attributes,
@@ -370,7 +370,7 @@ void SPIRVToOCL20::visitCallSPIRVAtomicBuiltin(CallInst* CI, Op OC) {
       // value by pointer passed as 2nd argument (aka expected) while SPIR-V
       // instructions returns this new/original value as a resulting value.
       AllocaInst *pExpected = new AllocaInst(CI->getType(), "expected",
-        pInsertBefore->getParent()->getParent()->getEntryBlock().getFirstInsertionPt());
+        &*pInsertBefore->getParent()->getParent()->getEntryBlock().getFirstInsertionPt());
       pExpected->setAlignment(CI->getType()->getScalarSizeInBits() / 8);
       new StoreInst(Args[1], pExpected, pInsertBefore);
       Args[1] = pExpected;
diff --git a/SPIRVUtil.cpp b/SPIRVUtil.cpp
index 7c5466b..a9fec78 100644
--- a/SPIRVUtil.cpp
+++ b/SPIRVUtil.cpp
@@ -45,7 +45,8 @@
 #include "OCLUtil.h"
 
 #include "llvm/ADT/StringSwitch.h"
-#include "llvm/Bitcode/ReaderWriter.h"
+#include "llvm/Bitcode/BitcodeWriter.h"
+#include "llvm/Bitcode/BitcodeReader.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
@@ -82,8 +83,7 @@ addFnAttr(LLVMContext *Context, CallInst *Call, Attribute::AttrKind Attr) {
 
 void
 removeFnAttr(LLVMContext *Context, CallInst *Call, Attribute::AttrKind Attr) {
-  Call->removeAttribute(AttributeSet::FunctionIndex,
-      Attribute::get(*Context, Attr));
+  Call->removeAttribute(AttributeSet::FunctionIndex, Attr);
 }
 
 Value *
@@ -814,7 +814,7 @@ addBlockBind(Module *M, Function *InvokeFunc, Value *BlkCtx, Value *CtxLen,
 
 IntegerType* getSizetType(Module *M) {
   return IntegerType::getIntNTy(M->getContext(),
-    M->getDataLayout()->getPointerSizeInBits(0));
+    M->getDataLayout().getPointerSizeInBits(0));
 }
 
 Type *
@@ -1031,7 +1031,7 @@ transTypeDesc(Type *Ty, const BuiltinArgTypeMangleInfo &Info) {
     if (Name.empty()) {
       std::ostringstream OS;
       OS << reinterpret_cast<size_t>(Ty);
-      Name = std::string("struct_") + OS.str();
+      Name = (std::string("struct_") + OS.str()).c_str();
     }
     return SPIR::RefParamType(new SPIR::UserDefinedType(Name));
   }
@@ -1343,7 +1343,7 @@ bool
 eraseUselessFunctions(Module *M) {
   bool changed = false;
   for (auto I = M->begin(), E = M->end(); I != E;)
-    changed |= eraseIfNoUse(I++);
+    changed |= eraseIfNoUse(&*I++);
   return changed;
 }
 
diff --git a/SPIRVWriter.cpp b/SPIRVWriter.cpp
index 6f5e6fd..2f8190b 100644
--- a/SPIRVWriter.cpp
+++ b/SPIRVWriter.cpp
@@ -1,1685 +1,1686 @@
-//===- SPIRVWriter.cpp - Converts LLVM to SPIR-V ----------------*- C++ -*-===//
-//
-//                     The LLVM/SPIR-V Translator
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-// Copyright (c) 2014 Advanced Micro Devices, Inc. All rights reserved.
-//
-// Permission is hereby granted, free of charge, to any person obtaining a
-// copy of this software and associated documentation files (the "Software"),
-// to deal with the Software without restriction, including without limitation
-// the rights to use, copy, modify, merge, publish, distribute, sublicense,
-// and/or sell copies of the Software, and to permit persons to whom the
-// Software is furnished to do so, subject to the following conditions:
-//
-// Redistributions of source code must retain the above copyright notice,
-// this list of conditions and the following disclaimers.
-// Redistributions in binary form must reproduce the above copyright notice,
-// this list of conditions and the following disclaimers in the documentation
-// and/or other materials provided with the distribution.
-// Neither the names of Advanced Micro Devices, Inc., nor the names of its
-// contributors may be used to endorse or promote products derived from this
-// Software without specific prior written permission.
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-// CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
-// THE SOFTWARE.
-//
-//===----------------------------------------------------------------------===//
-/// \file
-///
-/// This file implements conversion of LLVM intermediate language to SPIR-V
-/// binary.
-///
-//===----------------------------------------------------------------------===//
-
-#include "SPIRVModule.h"
-#include "SPIRVEnum.h"
-#include "SPIRVEntry.h"
-#include "SPIRVType.h"
-#include "SPIRVValue.h"
-#include "SPIRVFunction.h"
-#include "SPIRVBasicBlock.h"
-#include "SPIRVInstruction.h"
-#include "SPIRVExtInst.h"
-#include "SPIRVUtil.h"
-#include "SPIRVInternal.h"
-#include "SPIRVMDWalker.h"
-#include "OCLTypeToSPIRV.h"
-#include "OCLUtil.h"
-
-#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/SetVector.h"
-#include "llvm/ADT/StringSwitch.h"
-#include "llvm/ADT/Triple.h"
-#include "llvm/Bitcode/ReaderWriter.h"
-#include "llvm/IR/Constants.h"
-#include "llvm/IR/DerivedTypes.h"
-#include "llvm/IR/DebugInfo.h"
-#include "llvm/IR/Function.h"
-#include "llvm/IR/InstrTypes.h"
-#include "llvm/IR/Instructions.h"
-#include "llvm/IR/Module.h"
-#include "llvm/IR/Operator.h"
-#include "llvm/IR/Verifier.h"
-#include "llvm/Pass.h"
-#include "llvm/PassSupport.h"
-#include "llvm/PassManager.h"
-#include "llvm/Support/Casting.h"
-#include "llvm/Support/CommandLine.h"
-#include "llvm/Support/Debug.h"
-#include "llvm/Support/raw_ostream.h"
-#include "llvm/Support/ToolOutputFile.h"
-#include "llvm/Transforms/IPO.h"
-
-#include <iostream>
-#include <list>
-#include <memory>
-#include <set>
-#include <sstream>
-#include <vector>
-#include <functional>
-#include <cstdlib>
-
-#define DEBUG_TYPE "spirv"
-
-using namespace llvm;
-using namespace SPIRV;
-using namespace OCLUtil;
-
-namespace llvm {
-  FunctionPass *createPromoteMemoryToRegisterPass();
-}
-
-namespace SPIRV{
-
-cl::opt<bool> SPIRVMemToReg("spirv-mem2reg", cl::init(true),
-    cl::desc("LLVM/SPIR-V translation enable mem2reg"));
-
-
-static void
-foreachKernelArgMD(MDNode *MD, SPIRVFunction *BF,
-    std::function<void(const std::string& Str,
-        SPIRVFunctionParameter *BA)>Func) {
-  for (unsigned I = 1, E = MD->getNumOperands(); I != E; ++I) {
-    SPIRVFunctionParameter *BA = BF->getArgument(I-1);
-    Func(getMDOperandAsString(MD, I), BA);
-  }
-}
-
-/// Information for translating OCL builtin.
-struct OCLBuiltinSPIRVTransInfo {
-  std::string UniqName;
-  /// Postprocessor of operands
-  std::function<void(std::vector<SPIRVWord>&)> PostProc;
-  OCLBuiltinSPIRVTransInfo(){
-    PostProc = [](std::vector<SPIRVWord>&){};
-  }
-};
-
-class LLVMToSPIRVDbgTran {
-public:
-  LLVMToSPIRVDbgTran(Module *TM = nullptr, SPIRVModule *TBM = nullptr)
-  :BM(TBM), M(TM){
-  }
-
-  void setModule(Module *Mod) { M = Mod;}
-  void setSPIRVModule(SPIRVModule *SMod) { BM = SMod;}
-
-  void transDbgInfo(Value *V, SPIRVValue *BV) {
-    if (auto I = dyn_cast<Instruction>(V)) {
-      auto DL = I->getDebugLoc();
-      if (!DL.isUnknown()) {
-        DILocation DIL(DL.getAsMDNode());
-        auto File = BM->getString(DIL.getFilename().str());
-        // ToDo: SPIR-V rev.31 cannot add debug info for instructions without ids.
-        // This limitation needs to be addressed.
-        if (!BV->hasId())
-          return;
-        BM->addLine(BV, File, DIL.getLineNumber(), DIL.getColumnNumber());
-      }
-    } else if (auto F = dyn_cast<Function>(V)) {
-      if (auto DIS = getDISubprogram(F)) {
-        auto File = BM->getString(DIS.getFilename().str());
-        BM->addLine(BV, File, DIS.getLineNumber(), 0);
-      }
-    }
-  }
-
-private:
-  SPIRVModule *BM;
-  Module *M;
-};
-
-class LLVMToSPIRV: public ModulePass {
-public:
-  LLVMToSPIRV(SPIRVModule *SMod = nullptr)
-      : ModulePass(ID),
-        M(nullptr),
-        Ctx(nullptr),
-        BM(SMod),
-        ExtSetId(SPIRVID_INVALID),
-        SrcLang(0),
-        SrcLangVer(0),
-        DbgTran(nullptr, SMod){
-  }
-
-  bool runOnModule(Module &Mod) override {
-    M = &Mod;
-    Ctx = &M->getContext();
-    DbgTran.setModule(M);
-    assert(BM && "SPIR-V module not initialized");
-    translate();
-    return true;
-  }
-
-  void getAnalysisUsage(AnalysisUsage &AU) const {
-    AU.addRequired<OCLTypeToSPIRV>();
-  }
-
-  static char ID;
-
-  SPIRVType *transType(Type *T);
-  SPIRVType *transSPIRVOpaqueType(Type *T);
-
-  SPIRVValue *getTranslatedValue(Value *);
-
-  // Translation functions
-  bool transAddressingMode();
-  bool transAlign(Value *V, SPIRVValue *BV);
-  std::vector<SPIRVValue *> transArguments(CallInst *, SPIRVBasicBlock *);
-  std::vector<SPIRVWord> transArguments(CallInst *, SPIRVBasicBlock *,
-      SPIRVEntry *);
-  bool transSourceLanguage();
-  bool transExtension();
-  bool transBuiltinSet();
-  SPIRVValue *transCallInst(CallInst *Call, SPIRVBasicBlock *BB);
-  bool transDecoration(Value *V, SPIRVValue *BV);
-  SPIRVWord transFunctionControlMask(CallInst *);
-  SPIRVWord transFunctionControlMask(Function *);
-  SPIRVFunction *transFunctionDecl(Function *F);
-  bool transGlobalVariables();
-
-  Op transBoolOpCode(SPIRVValue *Opn, Op OC);
-  // Translate LLVM module to SPIR-V module.
-  // Returns true if succeeds.
-  bool translate();
-  bool transExecutionMode();
-  SPIRVValue *transConstant(Value *V);
-  SPIRVValue *transValue(Value *V, SPIRVBasicBlock *BB,
-      bool CreateForward = true);
-  SPIRVValue *transValueWithoutDecoration(Value *V, SPIRVBasicBlock *BB,
-      bool CreateForward = true);
-
-  typedef DenseMap<Type *, SPIRVType *> LLVMToSPIRVTypeMap;
-  typedef DenseMap<Value *, SPIRVValue *> LLVMToSPIRVValueMap;
-private:
-  Module *M;
-  LLVMContext *Ctx;
-  SPIRVModule *BM;
-  LLVMToSPIRVTypeMap TypeMap;
-  LLVMToSPIRVValueMap ValueMap;
-  //ToDo: support multiple builtin sets. Currently assume one builtin set.
-  SPIRVId ExtSetId;
-  SPIRVWord SrcLang;
-  SPIRVWord SrcLangVer;
-  LLVMToSPIRVDbgTran DbgTran;
-
-  SPIRVType *mapType(Type *T, SPIRVType *BT) {
-    TypeMap[T] = BT;
-    SPIRVDBG(dbgs() << "[mapType] " << *T << " => ";
-             spvdbgs() << *BT << '\n');
-    return BT;
-  }
-
-  SPIRVValue *mapValue(Value *V, SPIRVValue *BV) {
-    auto Loc = ValueMap.find(V);
-    if (Loc != ValueMap.end()) {
-      if (Loc->second == BV)
-        return BV;
-      assert (Loc->second->isForward() &&
-          "LLVM Value is mapped to different SPIRV Values");
-      auto Forward = static_cast<SPIRVForward *>(Loc->second);
-      BV->setId(Forward->getId());
-      BM->replaceForward(Forward, BV);
-    }
-    ValueMap[V] = BV;
-    SPIRVDBG(dbgs() << "[mapValue] " << *V << " => ";
-             spvdbgs() << *BV << "\n");
-    return BV;
-  }
-
-  SPIRVType *getSPIRVType(Type *T) {
-      return TypeMap[T];
-  }
-
-  SPIRVValue *getSPIRVValue(Value *V) {
-      return ValueMap[V];
-  }
-
-  SPIRVErrorLog &getErrorLog() {
-    return BM->getErrorLog();
-  }
-
-  llvm::IntegerType* getSizetType();
-  std::vector<SPIRVValue*> transValue(const std::vector<Value *> &Values,
-      SPIRVBasicBlock* BB);
-  std::vector<SPIRVWord> transValue(const std::vector<Value *> &Values,
-      SPIRVBasicBlock* BB, SPIRVEntry *Entry);
-
-  SPIRVInstruction* transBinaryInst(BinaryOperator* B, SPIRVBasicBlock* BB);
-  SPIRVInstruction* transCmpInst(CmpInst* Cmp, SPIRVBasicBlock* BB);
-
-  void dumpUsers(Value *V);
-
-  template<class ExtInstKind>
-  bool oclGetExtInstIndex(const std::string &MangledName,
-      const std::string& DemangledName, SPIRVWord* EntryPoint);
-  void oclGetMutatedArgumentTypesByBuiltin(llvm::FunctionType* FT,
-      std::map<unsigned, Type*>& ChangedType, Function* F);
-
-  bool isBuiltinTransToInst(Function *F);
-  bool isBuiltinTransToExtInst(Function *F,
-      SPIRVExtInstSetKind *BuiltinSet = nullptr,
-      SPIRVWord *EntryPoint = nullptr,
-      SmallVectorImpl<std::string> *Dec = nullptr);
-  bool oclIsKernel(Function *F);
-
-  bool transOCLKernelMetadata();
-
-  SPIRVInstruction *transBuiltinToInst(const std::string& DemangledName,
-      const std::string &MangledName, CallInst* CI, SPIRVBasicBlock* BB);
-  SPIRVInstruction *transBuiltinToInstWithoutDecoration(Op OC,
-      CallInst* CI, SPIRVBasicBlock* BB);
-  void mutateFuncArgType(const std::map<unsigned, Type*>& ChangedType,
-      Function* F);
-
-  SPIRVValue *transSpcvCast(CallInst* CI, SPIRVBasicBlock *BB);
-  SPIRVValue *oclTransSpvcCastSampler(CallInst* CI, SPIRVBasicBlock *BB);
-
-  SPIRV::SPIRVInstruction* transUnaryInst(UnaryInstruction* U,
-      SPIRVBasicBlock* BB);
-
-  /// Add a 32 bit integer constant.
-  /// \return Id of the constant.
-  SPIRVId addInt32(int);
-  void transFunction(Function *I);
-  SPIRV::SPIRVLinkageTypeKind transLinkageType(const GlobalValue* GV);
-};
-
-
-SPIRVValue *
-LLVMToSPIRV::getTranslatedValue(Value *V) {
-  LLVMToSPIRVValueMap::iterator Loc = ValueMap.find(V);
-  if (Loc != ValueMap.end())
-    return Loc->second;
-  return nullptr;
-}
-
-bool
-LLVMToSPIRV::oclIsKernel(Function *F) {
-  if (F->getCallingConv() == CallingConv::SPIR_KERNEL)
-    return true;
-  return false;
-}
-
-bool
-LLVMToSPIRV::isBuiltinTransToInst(Function *F) {
-  std::string DemangledName;
-  if (!oclIsBuiltin(F->getName(), &DemangledName) &&
-      !isDecoratedSPIRVFunc(F, &DemangledName))
-    return false;
-  SPIRVDBG(spvdbgs() << "CallInst: demangled name: " << DemangledName << '\n');
-  return getSPIRVFuncOC(DemangledName) != OpNop;
-}
-
-bool
-LLVMToSPIRV::isBuiltinTransToExtInst(Function *F,
-    SPIRVExtInstSetKind *ExtSet,
-    SPIRVWord *ExtOp,
-    SmallVectorImpl<std::string> *Dec) {
-  std::string OrigName = F->getName();
-  std::string DemangledName;
-  if (!oclIsBuiltin(OrigName, &DemangledName))
-    return false;
-  DEBUG(dbgs() << "[oclIsBuiltinTransToExtInst] CallInst: demangled name: "
-      << DemangledName << '\n');
-  StringRef S = DemangledName;
-  if (!S.startswith(kSPIRVName::Prefix))
-    return false;
-  S = S.drop_front(strlen(kSPIRVName::Prefix));
-  auto Loc = S.find(kSPIRVPostfix::Divider);
-  auto ExtSetName = S.substr(0, Loc);
-  SPIRVExtInstSetKind Set = SPIRVEIS_Count;
-  if (!SPIRVExtSetShortNameMap::rfind(ExtSetName, &Set))
-    return false;
-  assert(Set == BM->getBuiltinSet(ExtSetId) &&
-      "Invalid extended instruction set");
-  assert(Set == SPIRVEIS_OpenCL && "Unsupported extended instruction set");
-
-  auto ExtOpName = S.substr(Loc + 1);
-  auto Splited = ExtOpName.split(kSPIRVPostfix::ExtDivider);
-  OCLExtOpKind EOC;
-  if (!OCLExtOpMap::rfind(Splited.first, &EOC))
-    return false;
-
-  if (ExtSet)
-    *ExtSet = Set;
-  if (ExtOp)
-    *ExtOp = EOC;
-  if (Dec) {
-    SmallVector<StringRef, 2> P;
-    Splited.second.split(P, kSPIRVPostfix::Divider);
-    for (auto &I:P)
-      Dec->push_back(I.str());
-  }
-  return true;
-}
-
-/// Decode SPIR-V type name in the format spirv.{TypeName}._{Postfixes}
-/// where Postfixes are strings separated by underscores.
-/// \return TypeName.
-/// \param Ops contains the integers decoded from postfixes.
-static std::string
- decodeSPIRVTypeName(StringRef Name,
-    SmallVectorImpl<std::string>& Strs) {
-  SmallVector<StringRef, 4> SubStrs;
-  const char Delim[] = { kSPIRVTypeName::Delimiter, 0 };
-  Name.split(SubStrs, Delim, -1, true);
-  assert(SubStrs.size() >= 2 && "Invalid SPIRV type name");
-  assert(SubStrs[0] == kSPIRVTypeName::Prefix && "Invalid prefix");
-  assert((SubStrs.size() == 2 || !SubStrs[2].empty()) && "Invalid postfix");
-
-  if (SubStrs.size() > 2) {
-    const char PostDelim[] = { kSPIRVTypeName::PostfixDelim, 0 };
-    SmallVector<StringRef, 4> Postfixes;
-    SubStrs[2].split(Postfixes, PostDelim, -1, true);
-    assert(Postfixes.size() > 1 && Postfixes[0].empty() && "Invalid postfix");
-    for (unsigned I = 1, E = Postfixes.size(); I != E; ++I)
-      Strs.push_back(std::string(Postfixes[I]).c_str());
-  }
-  return SubStrs[1].str();
-}
-
-static bool recursiveType(const StructType *ST, const Type *Ty) {
-  SmallPtrSet<const StructType *, 4> Seen;
-
-  std::function<bool(const Type *Ty)> Run = [&](const Type *Ty) {
-    if (!isa<CompositeType>(Ty))
-      return false;
-
-    if (auto *StructTy = dyn_cast<StructType>(Ty)) {
-      if (StructTy == ST)
-        return true;
-
-      if (Seen.count(StructTy))
-        return false;
-
-      Seen.insert(StructTy);
-
-      return find_if(StructTy->subtype_begin(), StructTy->subtype_end(), Run) !=
-             StructTy->subtype_end();
-    }
-
-    if (auto *PtrTy = dyn_cast<PointerType>(Ty))
-      return Run(PtrTy->getPointerElementType());
-
-    if (auto *ArrayTy = dyn_cast<ArrayType>(Ty))
-      return Run(ArrayTy->getArrayElementType());
-
-    return false;
-  };
-
-  return Run(Ty);
-}
-
-SPIRVType *
-LLVMToSPIRV::transType(Type *T) {
-  LLVMToSPIRVTypeMap::iterator Loc = TypeMap.find(T);
-  if (Loc != TypeMap.end())
-    return Loc->second;
-
-  SPIRVDBG(dbgs() << "[transType] " << *T << '\n');
-  if (T->isVoidTy())
-    return mapType(T, BM->addVoidType());
-
-  if (T->isIntegerTy(1))
-    return mapType(T, BM->addBoolType());
-
-  if (T->isIntegerTy())
-    return mapType(T, BM->addIntegerType(T->getIntegerBitWidth()));
-
-  if (T->isFloatingPointTy())
-    return mapType(T, BM->addFloatType(T->getPrimitiveSizeInBits()));
-
-  // A pointer to image or pipe type in LLVM is translated to a SPIRV
-  // sampler or pipe type.
-  if (T->isPointerTy()) {
-    auto ET = T->getPointerElementType();
-    assert(!ET->isFunctionTy() && "Function pointer type is not allowed");
-    auto ST = dyn_cast<StructType>(ET);
-    auto AddrSpc = T->getPointerAddressSpace();
-    if (ST && !ST->isSized()) {
-      Op OpCode;
-      StringRef STName = ST->getName();
-      // Workaround for non-conformant SPIR binary
-      if (STName == "struct._event_t") {
-        STName = kSPR2TypeName::Event;
-        ST->setName(STName);
-      }
-      assert (!STName.startswith(kSPR2TypeName::Pipe) &&
-              "OpenCL type names should be translated to SPIR-V type names");
-      // ToDo: For SPIR1.2/2.0 there may still be load/store or bitcast
-      // instructions using opencl.* type names. We need to handle these
-      // type names until they are all mapped or FE generates SPIR-V type
-      // names.
-      if (STName.find(kSPR2TypeName::Pipe) == 0) {
-        assert(AddrSpc == SPIRAS_Global);
-        SmallVector<StringRef, 4> SubStrs;
-        const char Delims[] = {kSPR2TypeName::Delimiter, 0};
-        STName.split(SubStrs, Delims);
-        std::string Acc = kAccessQualName::ReadOnly;
-        if (SubStrs.size() > 2) {
-          Acc = SubStrs[2];
-        }
-        auto PipeT = BM->addPipeType();
-        PipeT->setPipeAcessQualifier(SPIRSPIRVAccessQualifierMap::map(Acc));
-        return mapType(T, PipeT);
-      } else if (STName.find(kSPR2TypeName::ImagePrefix) == 0) {
-        assert(AddrSpc == SPIRAS_Global);
-        auto SPIRVImageTy = getSPIRVImageTypeFromOCL(M, T);
-        return mapType(T, transSPIRVOpaqueType(SPIRVImageTy));
-      } else if (STName.startswith(kSPIRVTypeName::PrefixAndDelim))
-        return transSPIRVOpaqueType(T);
-      else if (OCLOpaqueTypeOpCodeMap::find(STName, &OpCode)) {
-        if (OpCode == OpTypePipe) {
-          return mapType(T, BM->addPipeType());
-        }
-        return mapType(T, BM->addOpaqueGenericType(OpCode));
-      } else if (isPointerToOpaqueStructType(T)) {
-        return mapType(T, BM->addPointerType(SPIRSPIRVAddrSpaceMap::map(
-          static_cast<SPIRAddressSpace>(AddrSpc)),
-          transType(ET)));
-      }
-    } else  {
-      return mapType(T, BM->addPointerType(SPIRSPIRVAddrSpaceMap::map(
-        static_cast<SPIRAddressSpace>(AddrSpc)),
-        transType(ET)));
-    }
-  }
-
-  if (T->isVectorTy())
-    return mapType(T, BM->addVectorType(transType(T->getVectorElementType()),
-        T->getVectorNumElements()));
-
-  if (T->isArrayTy())
-    return mapType(T, BM->addArrayType(transType(T->getArrayElementType()),
-        static_cast<SPIRVConstant*>(transValue(ConstantInt::get(getSizetType(),
-            T->getArrayNumElements(), false), nullptr))));
-
-  if (T->isStructTy() && !T->isSized()) {
-    auto ST = dyn_cast<StructType>(T);
-    assert(!ST->getName().startswith(kSPR2TypeName::Pipe));
-    assert(!ST->getName().startswith(kSPR2TypeName::ImagePrefix));
-    return mapType(T, BM->addOpaqueType(T->getStructName()));
-  }
-
-  if (auto ST = dyn_cast<StructType>(T)) {
-    assert(ST->isSized());
-
-    std::string Name;
-    if (ST->hasName())
-      Name = ST->getName();
-
-    if(Name == getSPIRVTypeName(kSPIRVTypeName::ConstantSampler))
-      return transType(getSamplerType(M));
-
-    auto *Struct = BM->openStructType(T->getStructNumElements(), Name);
-    mapType(T, Struct);
-
-    SmallVector<unsigned, 4> ForwardRefs;
-
-    for (unsigned I = 0, E = T->getStructNumElements(); I != E; ++I) {
-      auto *ElemTy = ST->getElementType(I);
-      if (isa<CompositeType>(ElemTy) && recursiveType(ST, ElemTy))
-        ForwardRefs.push_back(I);
-      else
-        Struct->setMemberType(I, transType(ST->getElementType(I)));
-    }
-
-    BM->closeStructType(Struct, ST->isPacked());
-
-    for (auto I : ForwardRefs)
-      Struct->setMemberType(I, transType(ST->getElementType(I)));
-
-    return Struct;
-  }
-
-  if (FunctionType *FT = dyn_cast<FunctionType>(T)) {
-    SPIRVType *RT = transType(FT->getReturnType());
-    std::vector<SPIRVType *> PT;
-    for (FunctionType::param_iterator I = FT->param_begin(),
-        E = FT->param_end(); I != E; ++I)
-      PT.push_back(transType(*I));
-    return mapType(T, BM->addFunctionType(RT, PT));
-  }
-
-  llvm_unreachable("Not implemented!");
-  return 0;
-}
-
-SPIRVType *
-LLVMToSPIRV::transSPIRVOpaqueType(Type *T) {
-  auto ET = T->getPointerElementType();
-  auto ST = cast<StructType>(ET);
-  auto AddrSpc = T->getPointerAddressSpace();
-  auto STName = ST->getStructName();
-  assert (STName.startswith(kSPIRVTypeName::PrefixAndDelim) &&
-    "Invalid SPIR-V opaque type name");
-  SmallVector<std::string, 8> Postfixes;
-  auto TN = decodeSPIRVTypeName(STName, Postfixes);
-  if (TN == kSPIRVTypeName::Pipe) {
-    assert(AddrSpc == SPIRAS_Global);
-    assert(Postfixes.size() == 1 && "Invalid pipe type ops");
-    auto PipeT = BM->addPipeType();
-    PipeT->setPipeAcessQualifier(static_cast<spv::AccessQualifier>(
-      atoi(Postfixes[0].c_str())));
-    return mapType(T, PipeT);
-  } else if (TN == kSPIRVTypeName::Image) {
-    assert(AddrSpc == SPIRAS_Global);
-    // The sampled type needs to be translated through LLVM type to guarantee
-    // uniqueness.
-    auto SampledT = transType(getLLVMTypeForSPIRVImageSampledTypePostfix(
-      Postfixes[0], *Ctx));
-    SmallVector<int, 7> Ops;
-    for (unsigned I = 1; I < 8; ++I)
-      Ops.push_back(atoi(Postfixes[I].c_str()));
-    SPIRVTypeImageDescriptor Desc(static_cast<SPIRVImageDimKind>(Ops[0]),
-        Ops[1], Ops[2], Ops[3], Ops[4], Ops[5]);
-    return mapType(T, BM->addImageType(SampledT, Desc,
-                   static_cast<spv::AccessQualifier>(Ops[6])));
-  } else if (TN == kSPIRVTypeName::SampledImg) {
-    return mapType(T, BM->addSampledImageType(
-        static_cast<SPIRVTypeImage *>(
-            transType(getSPIRVTypeByChangeBaseTypeName(M,
-                T, kSPIRVTypeName::SampledImg,
-                kSPIRVTypeName::Image)))));
-  } else if(TN == kSPIRVTypeName::Sampler)
-    return mapType(T, BM->addSamplerType());
-  else if (TN == kSPIRVTypeName::DeviceEvent)
-    return mapType(T, BM->addDeviceEventType());
-  else
-    return mapType(T, BM->addOpaqueGenericType(
-      SPIRVOpaqueTypeOpCodeMap::map(TN)));
-}
-
-SPIRVFunction *
-LLVMToSPIRV::transFunctionDecl(Function *F) {
-  if (auto BF= getTranslatedValue(F))
-    return static_cast<SPIRVFunction *>(BF);
-
-  SPIRVTypeFunction *BFT = static_cast<SPIRVTypeFunction *>(transType(
-      getAnalysis<OCLTypeToSPIRV>().getAdaptedType(F)));
-  SPIRVFunction *BF = static_cast<SPIRVFunction *>(mapValue(F,
-      BM->addFunction(BFT)));
-  BF->setFunctionControlMask(transFunctionControlMask(F));
-  if (F->hasName())
-    BM->setName(BF, F->getName());
-  if (oclIsKernel(F))
-    BM->addEntryPoint(ExecutionModelKernel, BF->getId());
-  else if (F->getLinkage() != GlobalValue::InternalLinkage)
-    BF->setLinkageType(transLinkageType(F));
-  auto Attrs = F->getAttributes();
-  for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
-      ++I) {
-    auto ArgNo = I->getArgNo();
-    SPIRVFunctionParameter *BA = BF->getArgument(ArgNo);
-    if (I->hasName())
-      BM->setName(BA, I->getName());
-    if (I->hasByValAttr())
-      BA->addAttr(FunctionParameterAttributeByVal);
-    if (I->hasNoAliasAttr())
-      BA->addAttr(FunctionParameterAttributeNoAlias);
-    if (I->hasNoCaptureAttr())
-      BA->addAttr(FunctionParameterAttributeNoCapture);
-    if (I->hasStructRetAttr())
-      BA->addAttr(FunctionParameterAttributeSret);
-    if (Attrs.hasAttribute(ArgNo + 1, Attribute::ZExt))
-      BA->addAttr(FunctionParameterAttributeZext);
-    if (Attrs.hasAttribute(ArgNo + 1, Attribute::SExt))
-      BA->addAttr(FunctionParameterAttributeSext);
-  }
-  if (Attrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt))
-    BF->addDecorate(DecorationFuncParamAttr, FunctionParameterAttributeZext);
-  if (Attrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
-    BF->addDecorate(DecorationFuncParamAttr, FunctionParameterAttributeSext);
-  DbgTran.transDbgInfo(F, BF);
-  SPIRVDBG(dbgs() << "[transFunction] " << *F << " => ";
-    spvdbgs() << *BF << '\n';)
-  return BF;
-}
-
-#define _SPIRV_OPL(x) OpLogical##x
-
-#define _SPIRV_OPB(x) OpBitwise##x
-
-SPIRVValue *
-LLVMToSPIRV::transConstant(Value *V) {
-  if (auto CPNull = dyn_cast<ConstantPointerNull>(V))
-    return BM->addNullConstant(bcast<SPIRVTypePointer>(transType(
-        CPNull->getType())));
-
-  if (auto CAZero = dyn_cast<ConstantAggregateZero>(V)) {
-    Type *AggType = CAZero->getType();
-    if (const StructType* ST = dyn_cast<StructType>(AggType)) {
-      if (ST->getName() == getSPIRVTypeName(kSPIRVTypeName::ConstantSampler))
-        return BM->addSamplerConstant(transType(AggType), 0,0,0);
-    } else
-      return BM->addNullConstant(transType(AggType));
-  }
-
-  if (auto ConstI = dyn_cast<ConstantInt>(V))
-    return BM->addConstant(transType(V->getType()), ConstI->getZExtValue());
-
-  if (auto ConstFP = dyn_cast<ConstantFP>(V)) {
-    auto BT = static_cast<SPIRVType *>(transType(V->getType()));
-    return BM->addConstant(BT,
-        ConstFP->getValueAPF().bitcastToAPInt().getZExtValue());
-  }
-
-  if (auto ConstDA = dyn_cast<ConstantDataArray>(V)) {
-    std::vector<SPIRVValue *> BV;
-    for (unsigned I = 0, E = ConstDA->getNumElements(); I != E; ++I)
-      BV.push_back(transValue(ConstDA->getElementAsConstant(I), nullptr));
-    return BM->addCompositeConstant(transType(V->getType()), BV);
-  }
-
-  if (auto ConstA = dyn_cast<ConstantArray>(V)) {
-    std::vector<SPIRVValue *> BV;
-    for (auto I = ConstA->op_begin(), E = ConstA->op_end(); I != E; ++I)
-      BV.push_back(transValue(*I, nullptr));
-    return BM->addCompositeConstant(transType(V->getType()), BV);
-  }
-
-  if (auto ConstDV = dyn_cast<ConstantDataVector>(V)) {
-    std::vector<SPIRVValue *> BV;
-    for (unsigned I = 0, E = ConstDV->getNumElements(); I != E; ++I)
-      BV.push_back(transValue(ConstDV->getElementAsConstant(I), nullptr));
-    return BM->addCompositeConstant(transType(V->getType()), BV);
-  }
-
-  if (auto ConstV = dyn_cast<ConstantVector>(V)) {
-    std::vector<SPIRVValue *> BV;
-    for (auto I = ConstV->op_begin(), E = ConstV->op_end(); I != E; ++I)
-      BV.push_back(transValue(*I, nullptr));
-    return BM->addCompositeConstant(transType(V->getType()), BV);
-  }
-
-  if (auto ConstV = dyn_cast<ConstantStruct>(V)) {
-    if (ConstV->getType()->getName() ==
-        getSPIRVTypeName(kSPIRVTypeName::ConstantSampler)) {
-      assert(ConstV->getNumOperands() == 3);
-      SPIRVWord
-        AddrMode   = ConstV->getOperand(0)->getUniqueInteger().getZExtValue(),
-        Normalized = ConstV->getOperand(1)->getUniqueInteger().getZExtValue(),
-        FilterMode = ConstV->getOperand(2)->getUniqueInteger().getZExtValue();
-      assert(AddrMode < 5 && "Invalid addressing mode");
-      assert(Normalized < 2 && "Invalid value of normalized coords");
-      assert(FilterMode < 2 && "Invalid filter mode");
-      SPIRVType* SamplerTy = transType(ConstV->getType());
-      return BM->addSamplerConstant(SamplerTy,
-                                    AddrMode, Normalized, FilterMode);
-    }
-    std::vector<SPIRVValue *> BV;
-    for (auto I = ConstV->op_begin(), E = ConstV->op_end(); I != E; ++I)
-      BV.push_back(transValue(*I, nullptr));
-    return BM->addCompositeConstant(transType(V->getType()), BV);
-  }
-
-  if (auto ConstUE = dyn_cast<ConstantExpr>(V)) {
-    auto Inst = ConstUE->getAsInstruction();
-    SPIRVDBG(dbgs() << "ConstantExpr: " << *ConstUE << '\n';
-      dbgs() << "Instruction: " << *Inst << '\n';)
-    auto BI = transValue(Inst, nullptr, false);
-    Inst->dropAllReferences();
-    return BI;
-  }
-
-  if (isa<UndefValue>(V)) {
-    return BM->addUndef(transType(V->getType()));
-  }
-
-  return nullptr;
-}
-
-SPIRVValue *
-LLVMToSPIRV::transValue(Value *V, SPIRVBasicBlock *BB, bool CreateForward) {
-  LLVMToSPIRVValueMap::iterator Loc = ValueMap.find(V);
-  if (Loc != ValueMap.end() && (!Loc->second->isForward() || CreateForward))
-    return Loc->second;
-
-  SPIRVDBG(dbgs() << "[transValue] " << *V << '\n');
-  assert ((!isa<Instruction>(V) || isa<GetElementPtrInst>(V) ||
-      isa<CastInst>(V) || BB) &&
-      "Invalid SPIRV BB");
-
-  auto BV = transValueWithoutDecoration(V, BB, CreateForward);
-  std::string name = V->getName();
-  if (!name.empty()) // Don't erase the name, which BM might already have
-    BM->setName(BV, name);
-  if(!transDecoration(V, BV))
-    return nullptr;
-  return BV;
-}
-
-SPIRVInstruction*
-LLVMToSPIRV::transBinaryInst(BinaryOperator* B, SPIRVBasicBlock* BB) {
-  unsigned LLVMOC = B->getOpcode();
-  auto Op0 = transValue(B->getOperand(0), BB);
-  SPIRVInstruction* BI = BM->addBinaryInst(
-      transBoolOpCode(Op0, OpCodeMap::map(LLVMOC)),
-      transType(B->getType()), Op0, transValue(B->getOperand(1), BB), BB);
-  return BI;
-}
-
-SPIRVInstruction*
-LLVMToSPIRV::transCmpInst(CmpInst* Cmp, SPIRVBasicBlock* BB) {
-  auto Op0 = transValue(Cmp->getOperand(0), BB);
-  SPIRVInstruction* BI = BM->addCmpInst(
-      transBoolOpCode(Op0, CmpMap::map(Cmp->getPredicate())),
-      transType(Cmp->getType()), Op0,
-      transValue(Cmp->getOperand(1), BB), BB);
-  return BI;
-}
-
-SPIRV::SPIRVInstruction *LLVMToSPIRV::transUnaryInst(UnaryInstruction *U,
-                                                  SPIRVBasicBlock *BB) {
-  Op BOC = OpNop;
-  if (auto Cast = dyn_cast<AddrSpaceCastInst>(U)) {
-    if (Cast->getDestTy()->getPointerAddressSpace() == SPIRAS_Generic) {
-      assert(Cast->getSrcTy()->getPointerAddressSpace() != SPIRAS_Constant &&
-             "Casts from constant address space to generic are illegal");
-      BOC = OpPtrCastToGeneric;
-    } else {
-      assert(Cast->getDestTy()->getPointerAddressSpace() != SPIRAS_Constant &&
-             "Casts from generic address space to constant are illegal");
-      assert(Cast->getSrcTy()->getPointerAddressSpace() == SPIRAS_Generic);
-      BOC = OpGenericCastToPtr;
-    }
-  } else {
-    auto OpCode = U->getOpcode();
-    BOC = OpCodeMap::map(OpCode);
-  }
-
-  auto Op = transValue(U->getOperand(0), BB);
-  return BM->addUnaryInst(transBoolOpCode(Op, BOC),
-      transType(U->getType()), Op, BB);
-}
-
-/// An instruction may use an instruction from another BB which has not been
-/// translated. SPIRVForward should be created as place holder for these
-/// instructions and replaced later by the real instructions.
-/// Use CreateForward = true to indicate such situation.
-SPIRVValue *
-LLVMToSPIRV::transValueWithoutDecoration(Value *V, SPIRVBasicBlock *BB,
-    bool CreateForward) {
-  if (auto LBB = dyn_cast<BasicBlock>(V)) {
-    auto BF = static_cast<SPIRVFunction *>(getTranslatedValue(LBB->getParent()));
-    assert (BF && "Function not translated");
-    BB = static_cast<SPIRVBasicBlock *>(mapValue(V, BM->addBasicBlock(BF)));
-    BM->setName(BB, LBB->getName());
-    return BB;
-  }
-
-  if (auto F = dyn_cast<Function>(V))
-    return transFunctionDecl(F);
-
-  if (auto GV = dyn_cast<GlobalVariable>(V)) {
-    llvm::PointerType * Ty = GV->getType();
-    // Though variables with common linkage type are initialized by 0,
-    // they can be represented in SPIR-V as uninitialized variables with
-    // 'Export' linkage type, just as tentative definitions look in C
-    llvm::Value *Init = GV->hasInitializer() && !GV->hasCommonLinkage() ?
-      GV->getInitializer() : nullptr;
-    StructType *ST = Init ? dyn_cast<StructType>(Init->getType()) : nullptr;
-    if (ST && ST->hasName() &&
-        ST->getName() == getSPIRVTypeName(kSPIRVTypeName::ConstantSampler)) {
-      auto BV = transConstant(Init);
-      assert(BV);
-      return mapValue(V, BV);
-    } else if (ConstantExpr *ConstUE = dyn_cast_or_null<ConstantExpr>(Init)) {
-      Instruction * Inst = ConstUE->getAsInstruction();
-      if (isSamplerInitializer(Inst)) {
-        Init = Inst->getOperand(0);
-        Ty = static_cast<PointerType*>(Init->getType());
-      }
-      Inst->dropAllReferences();
-    }
-    auto BVar = static_cast<SPIRVVariable *>(BM->addVariable(
-      transType(Ty), GV->isConstant(),
-      transLinkageType(GV),
-      Init ? transValue(Init, nullptr) : nullptr,
-      GV->getName(),
-      SPIRSPIRVAddrSpaceMap::map(
-        static_cast<SPIRAddressSpace>(Ty->getAddressSpace())),
-      nullptr
-      ));
-    mapValue(V, BVar);
-    spv::BuiltIn Builtin = spv::BuiltInPosition;
-    if (!GV->hasName() || !getSPIRVBuiltin(GV->getName().str(), Builtin))
-      return BVar;
-    BVar->setBuiltin(Builtin);
-    return BVar;
-  }
-
-  if (isa<Constant>(V)) {
-    auto BV = transConstant(V);
-    assert(BV);
-    return mapValue(V, BV);
-  }
-
-  if (auto Arg = dyn_cast<Argument>(V)) {
-    unsigned ArgNo = Arg->getArgNo();
-    SPIRVFunction *BF = BB->getParent();
-    //assert(BF->existArgument(ArgNo));
-    return mapValue(V, BF->getArgument(ArgNo));
-  }
-
-  if (CreateForward)
-    return mapValue(V, BM->addForward(transType(V->getType())));
-
-  if (StoreInst *ST = dyn_cast<StoreInst>(V)) {
-    std::vector<SPIRVWord> MemoryAccess;
-    if (ST->isVolatile())
-      MemoryAccess.push_back(MemoryAccessVolatileMask);
-    MemoryAccess.push_back(MemoryAccessAlignedMask);
-    MemoryAccess.push_back(ST->getAlignment());
-    return mapValue(V, BM->addStoreInst(
-        transValue(ST->getPointerOperand(), BB),
-        transValue(ST->getValueOperand(), BB),
-        MemoryAccess, BB));
-  }
-
-  if (LoadInst *LD = dyn_cast<LoadInst>(V)) {
-    std::vector<SPIRVWord> MemoryAccess;
-    if (LD->isVolatile())
-      MemoryAccess.push_back(MemoryAccessVolatileMask);
-    MemoryAccess.push_back(MemoryAccessAlignedMask);
-    MemoryAccess.push_back(LD->getAlignment());
-    return mapValue(V, BM->addLoadInst(
-        transValue(LD->getPointerOperand(), BB),
-        MemoryAccess, BB));
-  }
-
-  if (BinaryOperator *B = dyn_cast<BinaryOperator>(V)) {
-    SPIRVInstruction* BI = transBinaryInst(B, BB);
-    return mapValue(V, BI);
-  }
-
-  if (auto RI = dyn_cast<ReturnInst>(V)) {
-    if (auto RV = RI->getReturnValue())
-      return mapValue(V, BM->addReturnValueInst(
-          transValue(RV, BB), BB));
-    return mapValue(V, BM->addReturnInst(BB));
-  }
-
-  if (CmpInst *Cmp = dyn_cast<CmpInst>(V)) {
-    SPIRVInstruction* BI = transCmpInst(Cmp, BB);
-    return mapValue(V, BI);
-  }
-
-  if (SelectInst *Sel = dyn_cast<SelectInst>(V))
-    return mapValue(V, BM->addSelectInst(
-        transValue(Sel->getCondition(), BB),
-        transValue(Sel->getTrueValue(), BB),
-        transValue(Sel->getFalseValue(), BB),BB));
-
-  if (AllocaInst *Alc = dyn_cast<AllocaInst>(V))
-    return mapValue(V, BM->addVariable(
-      transType(Alc->getType()), false,
-      SPIRVLinkageTypeKind::LinkageTypeInternal,
-      nullptr, Alc->getName(),
-      StorageClassFunction, BB));
-
-  if (auto *Switch = dyn_cast<SwitchInst>(V)) {
-    std::vector<std::pair<SPIRVWord, SPIRVBasicBlock *>> Pairs;
-    for (auto I = Switch->case_begin(), E = Switch->case_end(); I != E; ++I)
-      Pairs.push_back(std::make_pair(I.getCaseValue()->getZExtValue(),
-          static_cast<SPIRVBasicBlock*>(transValue(I.getCaseSuccessor(),
-              nullptr))));
-    return mapValue(V, BM->addSwitchInst(
-        transValue(Switch->getCondition(), BB),
-        static_cast<SPIRVBasicBlock*>(transValue(Switch->getDefaultDest(),
-            nullptr)), Pairs, BB));
-  }
-
-  if (auto Branch = dyn_cast<BranchInst>(V)) {
-    if (Branch->isUnconditional())
-      return mapValue(V, BM->addBranchInst(
-          static_cast<SPIRVLabel*>(transValue(Branch->getSuccessor(0), BB)),
-          BB));
-    return mapValue(V, BM->addBranchConditionalInst(
-        transValue(Branch->getCondition(), BB),
-        static_cast<SPIRVLabel*>(transValue(Branch->getSuccessor(0), BB)),
-        static_cast<SPIRVLabel*>(transValue(Branch->getSuccessor(1), BB)),
-        BB));
-  }
-
-  if (auto Phi = dyn_cast<PHINode>(V)) {
-    std::vector<SPIRVValue *> IncomingPairs;
-    for (size_t I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) {
-      IncomingPairs.push_back(transValue(Phi->getIncomingValue(I), BB));
-      IncomingPairs.push_back(transValue(Phi->getIncomingBlock(I), nullptr));
-    }
-    return mapValue(V, BM->addPhiInst(transType(Phi->getType()), IncomingPairs,
-        BB));
-  }
-
-  if (UnaryInstruction *U = dyn_cast<UnaryInstruction>(V)) {
-    if(isSamplerInitializer(U))
-      return mapValue(V, transValue(U->getOperand(0), BB));
-    return mapValue(V, transUnaryInst(U, BB));
-  }
-
-  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
-    std::vector<SPIRVValue *> Indices;
-    for (unsigned i = 0, e = GEP->getNumIndices(); i != e; ++i)
-      Indices.push_back(transValue(GEP->getOperand(i+1), BB));
-    return mapValue(V, BM->addPtrAccessChainInst(
-        transType(GEP->getType()),
-        transValue(GEP->getPointerOperand(), BB),
-        Indices, BB, GEP->isInBounds()));
-  }
-
-  if (auto Ext = dyn_cast<ExtractElementInst>(V)) {
-    auto Index = Ext->getIndexOperand();
-    if (auto Const = dyn_cast<ConstantInt>(Index))
-      return mapValue(V, BM->addCompositeExtractInst(
-        transType(Ext->getType()),
-        transValue(Ext->getVectorOperand(), BB),
-        std::vector<SPIRVWord>(1, Const->getZExtValue()),
-        BB));
-    else
-      return mapValue(V, BM->addVectorExtractDynamicInst(
-          transValue(Ext->getVectorOperand(), BB),
-          transValue(Index, BB),
-          BB));
-  }
-
-  if (auto Ins = dyn_cast<InsertElementInst>(V)) {
-    auto Index = Ins->getOperand(2);
-    if (auto Const = dyn_cast<ConstantInt>(Index))
-      return mapValue(V, BM->addCompositeInsertInst(
-      transValue(Ins->getOperand(1), BB),
-      transValue(Ins->getOperand(0), BB),
-      std::vector<SPIRVWord>(1, Const->getZExtValue()),
-      BB));
-    else
-      return mapValue(V, BM->addVectorInsertDynamicInst(
-      transValue(Ins->getOperand(0), BB),
-      transValue(Ins->getOperand(1), BB),
-      transValue(Index, BB),
-      BB));
-  }
-
-  if (auto SF = dyn_cast<ShuffleVectorInst>(V)) {
-    std::vector<SPIRVWord> Comp;
-    for (auto &I:SF->getShuffleMask())
-      Comp.push_back(I);
-    return mapValue(V, BM->addVectorShuffleInst(
-        transType(SF->getType()),
-        transValue(SF->getOperand(0), BB),
-        transValue(SF->getOperand(1), BB),
-        Comp,
-        BB));
-  }
-
-  if (CallInst *CI = dyn_cast<CallInst>(V))
-    return mapValue(V, transCallInst(CI, BB));
-
-  llvm_unreachable("Not implemented");
-  return nullptr;
-}
-
-bool
-LLVMToSPIRV::transDecoration(Value *V, SPIRVValue *BV) {
-  if (!transAlign(V, BV))
-    return false;
-  if ((isa<AtomicCmpXchgInst>(V) &&
-      cast<AtomicCmpXchgInst>(V)->isVolatile()) ||
-      (isa<AtomicRMWInst>(V) && cast<AtomicRMWInst>(V)->isVolatile()))
-    BV->setVolatile(true);
-  DbgTran.transDbgInfo(V, BV);
-  return true;
-}
-
-bool
-LLVMToSPIRV::transAlign(Value *V, SPIRVValue *BV) {
-  if (auto AL = dyn_cast<AllocaInst>(V)) {
-    BM->setAlignment(BV, AL->getAlignment());
-    return true;
-  }
-  if (auto GV = dyn_cast<GlobalVariable>(V)) {
-    BM->setAlignment(BV, GV->getAlignment());
-    return true;
-  }
-  return true;
-}
-
-/// Do this after source language is set.
-bool
-LLVMToSPIRV::transBuiltinSet() {
-  SPIRVWord Ver = 0;
-  SourceLanguage Kind = BM->getSourceLanguage(&Ver);
-  assert((Kind == SourceLanguageOpenCL_C ||
-      Kind == SourceLanguageOpenCL_CPP ) && "not supported");
-  std::stringstream SS;
-  SS << "OpenCL.std";
-  return BM->importBuiltinSet(SS.str(), &ExtSetId);
-}
-
-/// Transform sampler* spcv.cast(i32 arg)
-/// Only two cases are possible:
-///   arg = ConstantInt x -> SPIRVConstantSampler
-///   arg = i32 argument -> transValue(arg)
-///   arg = load from sampler -> look through load
-SPIRVValue *
-LLVMToSPIRV::oclTransSpvcCastSampler(CallInst* CI, SPIRVBasicBlock *BB) {
-  llvm::Function* F = CI->getCalledFunction();
-  auto FT = F->getFunctionType();
-  auto RT = FT->getReturnType();
-  assert(FT->getNumParams() == 1);
-  assert(isSPIRVType(RT, kSPIRVTypeName::Sampler) &&
-    FT->getParamType(0)->isIntegerTy() && "Invalid sampler type");
-  auto Arg = CI->getArgOperand(0);
-
-  auto GetSamplerConstant = [&](uint64_t SamplerValue) {
-    auto AddrMode = (SamplerValue & 0xE) >> 1;
-    auto Param = SamplerValue & 0x1;
-    auto Filter = ((SamplerValue & 0x30) >> 4) - 1;
-    auto BV = BM->addSamplerConstant(transType(RT), AddrMode, Param, Filter);
-    return BV;
-  };
-
-  if (auto Const = dyn_cast<ConstantInt>(Arg)) {
-    // Sampler is declared as a kernel scope constant
-    return GetSamplerConstant(Const->getZExtValue());
-  } else if (auto Load = dyn_cast<LoadInst>(Arg)) {
-    // If value of the sampler is loaded from a global constant, use its
-    // initializer for initialization of the sampler.
-    auto Op = Load->getPointerOperand();
-    assert(isa<GlobalVariable>(Op) && "Unknown sampler pattern!");
-    auto GV = cast<GlobalVariable>(Op);
-    assert(GV->isConstant() ||
-      GV->getType()->getPointerAddressSpace() == SPIRAS_Constant);
-    auto Initializer = GV->getInitializer();
-    assert(isa<ConstantInt>(Initializer) && "sampler not constant int?");
-    return GetSamplerConstant(cast<ConstantInt>(Initializer)->getZExtValue());
-  }
-  // Sampler is a function argument
-  auto BV = transValue(Arg, BB);
-  assert(BV && BV->getType() == transType(RT));
-  return BV;
-}
-
-SPIRVValue *
-LLVMToSPIRV::transSpcvCast(CallInst* CI, SPIRVBasicBlock *BB) {
-  return oclTransSpvcCastSampler(CI, BB);
-}
-
-SPIRVValue *
-LLVMToSPIRV::transCallInst(CallInst *CI, SPIRVBasicBlock *BB) {
-  SPIRVExtInstSetKind ExtSetKind = SPIRVEIS_Count;
-  SPIRVWord ExtOp = SPIRVWORD_MAX;
-  llvm::Function* F = CI->getCalledFunction();
-  auto MangledName = F->getName();
-  std::string DemangledName;
-
-  if (MangledName.startswith(SPCV_CAST))
-    return transSpcvCast(CI, BB);
-
-  if (MangledName.startswith("llvm.memcpy")) {
-    std::vector<SPIRVWord> MemoryAccess;
-
-    if (isa<ConstantInt>(CI->getOperand(4)) &&
-      dyn_cast<ConstantInt>(CI->getOperand(4))
-      ->getZExtValue() == 1)
-      MemoryAccess.push_back(MemoryAccessVolatileMask);
-    if (isa<ConstantInt>(CI->getOperand(3))) {
-        MemoryAccess.push_back(MemoryAccessAlignedMask);
-        MemoryAccess.push_back(dyn_cast<ConstantInt>(CI->getOperand(3))
-          ->getZExtValue());
-    }
-
-    return BM->addCopyMemorySizedInst(
-      transValue(CI->getOperand(0), BB),
-      transValue(CI->getOperand(1), BB),
-      transValue(CI->getOperand(2), BB),
-      MemoryAccess,
-      BB);
-  }
-
-  if (oclIsBuiltin(MangledName, &DemangledName) ||
-      isDecoratedSPIRVFunc(F, &DemangledName))
-    if (auto BV = transBuiltinToInst(DemangledName, MangledName, CI, BB))
-      return BV;
-
-  SmallVector<std::string, 2> Dec;
-  if (isBuiltinTransToExtInst(CI->getCalledFunction(), &ExtSetKind,
-      &ExtOp, &Dec))
-    return addDecorations(BM->addExtInst(
-      transType(CI->getType()),
-      ExtSetId,
-      ExtOp,
-      transArguments(CI, BB, SPIRVEntry::create_unique(ExtSetKind, ExtOp).get()),
-      BB), Dec);
-
-  return BM->addCallInst(
-    transFunctionDecl(CI->getCalledFunction()),
-    transArguments(CI, BB, SPIRVEntry::create_unique(OpFunctionCall).get()),
-    BB);
-}
-
-bool
-LLVMToSPIRV::transAddressingMode() {
-  Triple TargetTriple(M->getTargetTriple());
-  Triple::ArchType Arch = TargetTriple.getArch();
-
-  SPIRVCKRT(Arch == Triple::spir || Arch == Triple::spir64,
-      InvalidTargetTriple,
-      "Actual target triple is " + M->getTargetTriple());
-
-  if (Arch == Triple::spir)
-    BM->setAddressingModel(AddressingModelPhysical32);
-  else
-    BM->setAddressingModel(AddressingModelPhysical64);
-  // Physical addressing model requires Addresses capability
-  BM->addCapability(CapabilityAddresses);
-  return true;
-}
-std::vector<SPIRVValue*>
-LLVMToSPIRV::transValue(const std::vector<Value *> &Args, SPIRVBasicBlock* BB) {
-  std::vector<SPIRVValue*> BArgs;
-  for (auto &I: Args)
-    BArgs.push_back(transValue(I, BB));
-  return BArgs;
-}
-
-std::vector<SPIRVValue*>
-LLVMToSPIRV::transArguments(CallInst *CI, SPIRVBasicBlock *BB) {
-  return transValue(getArguments(CI), BB);
-}
-
-std::vector<SPIRVWord>
-LLVMToSPIRV::transValue(const std::vector<Value *> &Args, SPIRVBasicBlock* BB,
-    SPIRVEntry *Entry) {
-  std::vector<SPIRVWord> Operands;
-  for (size_t I = 0, E = Args.size(); I != E; ++I) {
-    Operands.push_back(Entry->isOperandLiteral(I) ?
-        cast<ConstantInt>(Args[I])->getZExtValue() :
-        transValue(Args[I], BB)->getId());
-  }
-  return Operands;
-}
-
-std::vector<SPIRVWord>
-LLVMToSPIRV::transArguments(CallInst *CI, SPIRVBasicBlock *BB, SPIRVEntry *Entry) {
-  return transValue(getArguments(CI), BB, Entry);
-}
-
-SPIRVWord
-LLVMToSPIRV::transFunctionControlMask(CallInst *CI) {
-  SPIRVWord FCM = 0;
-  SPIRSPIRVFuncCtlMaskMap::foreach([&](Attribute::AttrKind Attr,
-      SPIRVFunctionControlMaskKind Mask){
-    if (CI->hasFnAttr(Attr))
-      FCM |= Mask;
-  });
-  return FCM;
-}
-
-SPIRVWord
-LLVMToSPIRV::transFunctionControlMask(Function *F) {
-  SPIRVWord FCM = 0;
-  SPIRSPIRVFuncCtlMaskMap::foreach([&](Attribute::AttrKind Attr,
-      SPIRVFunctionControlMaskKind Mask){
-    if (F->hasFnAttribute(Attr))
-      FCM |= Mask;
-  });
-  return FCM;
-}
-
-bool
-LLVMToSPIRV::transGlobalVariables() {
-  for (auto I = M->global_begin(),
-            E = M->global_end(); I != E; ++I) {
-    if (!transValue(I, nullptr))
-      return false;
-  }
-  return true;
-}
-
-void
-LLVMToSPIRV::mutateFuncArgType(const std::map<unsigned, Type*>& ChangedType,
-    Function* F) {
-  for (auto &I : ChangedType) {
-    for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE; ++UI) {
-      auto Call = dyn_cast<CallInst>(*UI);
-      if (!Call)
-        continue;
-      auto Arg = Call->getArgOperand(I.first);
-      auto OrigTy = Arg->getType();
-      if (OrigTy == I.second)
-        continue;
-      SPIRVDBG(dbgs() << "[mutate arg type] " << *Call << ", " << *Arg << '\n');
-      auto CastF = M->getOrInsertFunction(SPCV_CAST, I.second, OrigTy, nullptr);
-      std::vector<Value *> Args;
-      Args.push_back(Arg);
-      auto Cast = CallInst::Create(CastF, Args, "", Call);
-      Call->replaceUsesOfWith(Arg, Cast);
-      SPIRVDBG(dbgs() << "[mutate arg type] -> " << *Cast << '\n');
-    }
-  }
-}
-
-void
-LLVMToSPIRV::transFunction(Function *I) {
-  transFunctionDecl(I);
-  // Creating all basic blocks before creating any instruction.
-  for (Function::iterator FI = I->begin(), FE = I->end(); FI != FE; ++FI) {
-    transValue(FI, nullptr);
-  }
-  for (Function::iterator FI = I->begin(), FE = I->end(); FI != FE; ++FI) {
-    SPIRVBasicBlock* BB = static_cast<SPIRVBasicBlock*>(transValue(FI, nullptr));
-    for (BasicBlock::iterator BI = FI->begin(), BE = FI->end(); BI != BE;
-        ++BI) {
-      transValue(BI, BB, false);
-    }
-  }
-}
-
-bool
-LLVMToSPIRV::translate() {
-  BM->setGeneratorVer(kTranslatorVer);
-
-  if (!transSourceLanguage())
-    return false;
-  if (!transExtension())
-    return false;
-  if (!transBuiltinSet())
-    return false;
-  if (!transAddressingMode())
-    return false;
-  if (!transGlobalVariables())
-    return false;
-
-  for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
-    Function *F = I;
-    auto FT = F->getFunctionType();
-    std::map<unsigned, Type *> ChangedType;
-    oclGetMutatedArgumentTypesByBuiltin(FT, ChangedType, F);
-    mutateFuncArgType(ChangedType, F);
-  }
-
-  // SPIR-V logical layout requires all function declarations go before
-  // function definitions.
-  std::vector<Function *> Decls, Defs;
-  for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
-    if (isBuiltinTransToInst(I) || isBuiltinTransToExtInst(I)
-        || I->getName().startswith(SPCV_CAST) ||
-        I->getName().startswith(LLVM_MEMCPY))
-      continue;
-    if (I->isDeclaration())
-      Decls.push_back(I);
-    else
-      Defs.push_back(I);
-  }
-  for (auto I:Decls)
-    transFunctionDecl(I);
-  for (auto I:Defs)
-    transFunction(I);
-
-  if (!transOCLKernelMetadata())
-    return false;
-  if (!transExecutionMode())
-    return false;
-
-  BM->optimizeDecorates();
-  BM->resolveUnknownStructFields();
-  BM->createForwardPointers();
-  return true;
-}
-
-llvm::IntegerType* LLVMToSPIRV::getSizetType() {
-  return IntegerType::getIntNTy(M->getContext(),
-    M->getDataLayout()->getPointerSizeInBits());
-}
-
-void
-LLVMToSPIRV::oclGetMutatedArgumentTypesByBuiltin(
-    llvm::FunctionType* FT, std::map<unsigned, Type*>& ChangedType,
-    Function* F) {
-  auto Name = F->getName();
-  std::string Demangled;
-  if (!oclIsBuiltin(Name, &Demangled))
-    return;
-  if (Demangled.find(kSPIRVName::SampledImage) == std::string::npos)
-    return;
-  if (FT->getParamType(1)->isIntegerTy())
-    ChangedType[1] = getSamplerType(F->getParent());
-}
-
-SPIRVInstruction *
-LLVMToSPIRV::transBuiltinToInst(const std::string& DemangledName,
-    const std::string &MangledName, CallInst* CI, SPIRVBasicBlock* BB) {
-  SmallVector<std::string, 2> Dec;
-  auto OC = getSPIRVFuncOC(DemangledName, &Dec);
-
-  if (OC == OpNop)
-    return nullptr;
-
-  auto Inst = transBuiltinToInstWithoutDecoration(OC, CI, BB);
-  addDecorations(Inst, Dec);
-  return Inst;
-}
-
-bool
-LLVMToSPIRV::transExecutionMode() {
-  if (auto NMD = SPIRVMDWalker(*M).getNamedMD(kSPIRVMD::ExecutionMode)) {
-    while (!NMD.atEnd()) {
-      unsigned EMode = ~0U;
-      unsigned EModel = ~0U;
-      Function *F = nullptr;
-      auto N = NMD.nextOp(); /* execution mode MDNode */
-      N.nextOp() /* entry point MDNode */
-          .get(EModel)
-          .get(F)
-          .done()
-       .get(EMode);
-      assert (EModel == spv::ExecutionModelKernel &&
-          "Unsupported execution model");
-      SPIRVFunction *BF = static_cast<SPIRVFunction *>(getTranslatedValue(F));
-      assert(BF && "Invalid kernel function");
-      switch(EMode) {
-      case spv::ExecutionModeContractionOff:
-        BF->addExecutionMode(new SPIRVExecutionMode(BF,
-            ExecutionModeContractionOff));
-        break;
-      case spv::ExecutionModeLocalSize: {
-        unsigned X, Y, Z;
-        N.get(X).get(Y).get(Z);
-        BF->addExecutionMode(new SPIRVExecutionMode(BF,
-            ExecutionModeLocalSize, X, Y, Z));
-      }
-      break;
-      case spv::ExecutionModeLocalSizeHint: {
-        unsigned X, Y, Z;
-        N.get(X).get(Y).get(Z);
-        BF->addExecutionMode(new SPIRVExecutionMode(BF,
-            ExecutionModeLocalSizeHint, X, Y, Z));
-      }
-      break;
-      case spv::ExecutionModeVecTypeHint: {
-        unsigned X;
-        N.get(X);
-        BF->addExecutionMode(new SPIRVExecutionMode(BF,
-            ExecutionModeVecTypeHint, X));
-      }
-      break;
-      default:
-        llvm_unreachable("invalid execution mode");
-      }
-    }
-  }
-  return true;
-}
-
-bool
-LLVMToSPIRV::transOCLKernelMetadata() {
-  NamedMDNode *KernelMDs = M->getNamedMetadata(SPIR_MD_KERNELS);
-  std::vector<std::string> argAccessQual;
-  if (!KernelMDs)
-    return true;
-
-  for (unsigned I = 0, E = KernelMDs->getNumOperands(); I < E; ++I) {
-    MDNode *KernelMD = KernelMDs->getOperand(I);
-    if (KernelMD->getNumOperands() == 0)
-      continue;
-    Function *Kernel = mdconst::dyn_extract<Function>(KernelMD->getOperand(0));
-
-    SPIRVFunction *BF = static_cast<SPIRVFunction *>(getTranslatedValue(Kernel));
-    assert(BF && "Kernel function should be translated first");
-    assert(Kernel && oclIsKernel(Kernel)
-            && "Invalid kernel calling convention or metadata");
-    for (unsigned MI = 1, ME = KernelMD->getNumOperands(); MI < ME; ++MI) {
-      MDNode *MD = dyn_cast<MDNode>(KernelMD->getOperand(MI));
-      if (!MD)
-        continue;
-      MDString *NameMD = dyn_cast<MDString>(MD->getOperand(0));
-      if (!NameMD)
-        continue;
-      StringRef Name = NameMD->getString();
-      if (Name == SPIR_MD_KERNEL_ARG_TYPE_QUAL) {
-        foreachKernelArgMD(MD, BF,
-            [](const std::string &Str, SPIRVFunctionParameter *BA){
-          if (Str.find("volatile") != std::string::npos)
-            BA->addDecorate(new SPIRVDecorate(DecorationVolatile, BA));
-          if (Str.find("restrict") != std::string::npos)
-            BA->addDecorate(new SPIRVDecorate(DecorationFuncParamAttr,
-                BA, FunctionParameterAttributeNoAlias));
-          if (Str.find("const") != std::string::npos)
-            BA->addDecorate(new SPIRVDecorate(DecorationFuncParamAttr,
-                BA, FunctionParameterAttributeNoWrite));
-          });
-      } else if (Name == SPIR_MD_KERNEL_ARG_NAME) {
-        foreachKernelArgMD(MD, BF,
-            [=](const std::string &Str, SPIRVFunctionParameter *BA){
-            BM->setName(BA, Str);
-          });
-      }
-    }
-  }
-  return true;
-}
-
-bool
-LLVMToSPIRV::transSourceLanguage() {
-  auto Src = getSPIRVSource(M);
-  SrcLang = std::get<0>(Src);
-  SrcLangVer = std::get<1>(Src);
-  BM->setSourceLanguage(static_cast<SourceLanguage>(SrcLang), SrcLangVer);
-  return true;
-}
-
-bool
-LLVMToSPIRV::transExtension() {
-  if (auto N = SPIRVMDWalker(*M).getNamedMD(kSPIRVMD::Extension)) {
-    while (!N.atEnd()) {
-      std::string S;
-      N.nextOp().get(S);
-      assert(!S.empty() && "Invalid extension");
-      BM->getExtension().insert(S);
-    }
-  }
-  if (auto N = SPIRVMDWalker(*M).getNamedMD(kSPIRVMD::SourceExtension)) {
-    while (!N.atEnd()) {
-      std::string S;
-      N.nextOp().get(S);
-      assert(!S.empty() && "Invalid extension");
-      BM->getSourceExtension().insert(S);
-    }
-  }
-  for (auto &I:map<SPIRVCapabilityKind>(rmap<OclExt::Kind>(BM->getExtension())))
-    BM->addCapability(I);
-
-  return true;
-}
-
-void
-LLVMToSPIRV::dumpUsers(Value* V) {
-  SPIRVDBG(dbgs() << "Users of " << *V << " :\n");
-  for (auto UI = V->user_begin(), UE = V->user_end();
-      UI != UE; ++UI)
-    SPIRVDBG(dbgs() << "  " << **UI << '\n');
-}
-
-Op
-LLVMToSPIRV::transBoolOpCode(SPIRVValue* Opn, Op OC) {
-  if (!Opn->getType()->isTypeVectorOrScalarBool())
-    return OC;
-  IntBoolOpMap::find(OC, &OC);
-  return OC;
-}
-
-SPIRVInstruction *
-LLVMToSPIRV::transBuiltinToInstWithoutDecoration(Op OC,
-    CallInst* CI, SPIRVBasicBlock* BB) {
-  if (isGroupOpCode(OC))
-    BM->addCapability(CapabilityGroups);
-  switch (OC) {
-  case OpControlBarrier: {
-    auto BArgs = transValue(getArguments(CI), BB);
-    return BM->addControlBarrierInst(
-      BArgs[0], BArgs[1], BArgs[2], BB);
-    }
-    break;
-  case OpGroupAsyncCopy: {
-    auto BArgs = transValue(getArguments(CI), BB);
-    return BM->addAsyncGroupCopy(BArgs[0], BArgs[1], BArgs[2], BArgs[3],
-                                 BArgs[4], BArgs[5], BB);
-    }
-    break;
-  default: {
-    if (isCvtOpCode(OC) && OC != OpGenericCastToPtrExplicit) {
-      return BM->addUnaryInst(OC, transType(CI->getType()),
-        transValue(CI->getArgOperand(0), BB), BB);
-    } else if (isCmpOpCode(OC)) {
-      assert(CI && CI->getNumArgOperands() == 2 && "Invalid call inst");
-      auto ResultTy = CI->getType();
-      Type *BoolTy = IntegerType::getInt1Ty(M->getContext());
-      auto IsVector = ResultTy->isVectorTy();
-      if (IsVector)
-        BoolTy = VectorType::get(BoolTy, ResultTy->getVectorNumElements());
-      auto BBT = transType(BoolTy);
-      auto Cmp = BM->addCmpInst(OC, BBT,
-        transValue(CI->getArgOperand(0), BB),
-        transValue(CI->getArgOperand(1), BB), BB);
-      auto Zero = transValue(Constant::getNullValue(ResultTy), BB);
-      auto One = transValue(
-          IsVector ? Constant::getAllOnesValue(ResultTy) : getInt32(M, 1), BB);
-      return BM->addSelectInst(Cmp, One, Zero, BB);
-    } else if (isBinaryOpCode(OC)) {
-      assert(CI && CI->getNumArgOperands() == 2 && "Invalid call inst");
-      return BM->addBinaryInst(OC, transType(CI->getType()),
-        transValue(CI->getArgOperand(0), BB),
-        transValue(CI->getArgOperand(1), BB), BB);
-    } else if (CI->getNumArgOperands() == 1 &&
-        !CI->getType()->isVoidTy() &&
-        !hasExecScope(OC) &&
-        !isAtomicOpCode(OC)) {
-      return BM->addUnaryInst(OC, transType(CI->getType()),
-        transValue(CI->getArgOperand(0), BB), BB);
-    } else {
-      auto Args = getArguments(CI);
-      SPIRVType *SPRetTy = nullptr;
-      Type *RetTy = CI->getType();
-      auto F = CI->getCalledFunction();
-      if (!RetTy->isVoidTy()) {
-        SPRetTy = transType(RetTy);
-      } else if (Args.size() > 0 && F->arg_begin()->hasStructRetAttr()) {
-        SPRetTy = transType(F->arg_begin()->getType()->getPointerElementType());
-        Args.erase(Args.begin());
-      }
-      auto SPI = BM->addInstTemplate(OC, BB, SPRetTy);
-      std::vector<SPIRVWord> SPArgs;
-      for (size_t I = 0, E = Args.size(); I != E; ++I) {
-        assert((!isFunctionPointerType(Args[I]->getType()) ||
-               isa<Function>(Args[I])) &&
-               "Invalid function pointer argument");
-        SPArgs.push_back(SPI->isOperandLiteral(I) ?
-            cast<ConstantInt>(Args[I])->getZExtValue() :
-            transValue(Args[I], BB)->getId());
-      }
-      SPI->setOpWordsAndValidate(SPArgs);
-      if (!SPRetTy || !SPRetTy->isTypeStruct())
-        return SPI;
-      std::vector<SPIRVWord> Mem;
-      SPIRVDBG(spvdbgs() << *SPI << '\n');
-      return BM->addStoreInst(transValue(CI->getArgOperand(0), BB), SPI,
-          Mem, BB);
-    }
-  }
-  }
-  return nullptr;
-}
-
-
-SPIRVId
-LLVMToSPIRV::addInt32(int I) {
-  return transValue(getInt32(M, I), nullptr, false)->getId();
-}
-
-SPIRV::SPIRVLinkageTypeKind
-LLVMToSPIRV::transLinkageType(const GlobalValue* GV) {
-  if(GV->isDeclarationForLinker())
-    return SPIRVLinkageTypeKind::LinkageTypeImport;
-  if(GV->hasInternalLinkage() || GV->hasPrivateLinkage())
-    return SPIRVLinkageTypeKind::LinkageTypeInternal;
-  return SPIRVLinkageTypeKind::LinkageTypeExport;
-}
-} // end of SPIRV namespace
-
-char LLVMToSPIRV::ID = 0;
-
-INITIALIZE_PASS_BEGIN(LLVMToSPIRV, "llvmtospv", "Translate LLVM to SPIR-V",
-    false, false)
-INITIALIZE_PASS_DEPENDENCY(OCLTypeToSPIRV)
-INITIALIZE_PASS_END(LLVMToSPIRV, "llvmtospv", "Translate LLVM to SPIR-V",
-    false, false)
-
-ModulePass *llvm::createLLVMToSPIRV(SPIRVModule *SMod) {
-  return new LLVMToSPIRV(SMod);
-}
-
-void
-addPassesForSPIRV(PassManager &PassMgr) {
-  if (SPIRVMemToReg)
-    PassMgr.add(createPromoteMemoryToRegisterPass());
-  PassMgr.add(createTransOCLMD());
-  PassMgr.add(createOCL21ToSPIRV());
-  PassMgr.add(createSPIRVLowerOCLBlocks());
-  PassMgr.add(createOCLTypeToSPIRV());
-  PassMgr.add(createOCL20ToSPIRV());
-  PassMgr.add(createSPIRVRegularizeLLVM());
-  PassMgr.add(createSPIRVLowerConstExpr());
-  PassMgr.add(createSPIRVLowerBool());
-}
-
-bool
-llvm::WriteSPIRV(Module *M, llvm::raw_ostream &OS, std::string &ErrMsg) {
-  std::unique_ptr<SPIRVModule> BM(SPIRVModule::createSPIRVModule());
-  PassManager PassMgr;
-  addPassesForSPIRV(PassMgr);
-  PassMgr.add(createLLVMToSPIRV(BM.get()));
-  PassMgr.run(*M);
-
-  if (BM->getError(ErrMsg) != SPIRVEC_Success)
-    return false;
-  OS << *BM;
-  return true;
-}
-
-bool
-llvm::RegularizeLLVMForSPIRV(Module *M, std::string &ErrMsg) {
-  std::unique_ptr<SPIRVModule> BM(SPIRVModule::createSPIRVModule());
-  PassManager PassMgr;
-  addPassesForSPIRV(PassMgr);
-  PassMgr.run(*M);
-  return true;
-}
-
+//===- SPIRVWriter.cpp - Converts LLVM to SPIR-V ----------------*- C++ -*-===//
+//
+//                     The LLVM/SPIR-V Translator
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+// Copyright (c) 2014 Advanced Micro Devices, Inc. All rights reserved.
+//
+// Permission is hereby granted, free of charge, to any person obtaining a
+// copy of this software and associated documentation files (the "Software"),
+// to deal with the Software without restriction, including without limitation
+// the rights to use, copy, modify, merge, publish, distribute, sublicense,
+// and/or sell copies of the Software, and to permit persons to whom the
+// Software is furnished to do so, subject to the following conditions:
+//
+// Redistributions of source code must retain the above copyright notice,
+// this list of conditions and the following disclaimers.
+// Redistributions in binary form must reproduce the above copyright notice,
+// this list of conditions and the following disclaimers in the documentation
+// and/or other materials provided with the distribution.
+// Neither the names of Advanced Micro Devices, Inc., nor the names of its
+// contributors may be used to endorse or promote products derived from this
+// Software without specific prior written permission.
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+// CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+// THE SOFTWARE.
+//
+//===----------------------------------------------------------------------===//
+/// \file
+///
+/// This file implements conversion of LLVM intermediate language to SPIR-V
+/// binary.
+///
+//===----------------------------------------------------------------------===//
+
+#include "SPIRVModule.h"
+#include "SPIRVEnum.h"
+#include "SPIRVEntry.h"
+#include "SPIRVType.h"
+#include "SPIRVValue.h"
+#include "SPIRVFunction.h"
+#include "SPIRVBasicBlock.h"
+#include "SPIRVInstruction.h"
+#include "SPIRVExtInst.h"
+#include "SPIRVUtil.h"
+#include "SPIRVInternal.h"
+#include "SPIRVMDWalker.h"
+#include "OCLTypeToSPIRV.h"
+#include "OCLUtil.h"
+
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SetVector.h"
+#include "llvm/ADT/StringSwitch.h"
+#include "llvm/ADT/Triple.h"
+#include "llvm/Bitcode/BitcodeWriter.h"
+#include "llvm/Bitcode/BitcodeReader.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/DerivedTypes.h"
+#include "llvm/IR/DebugInfo.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/InstrTypes.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/Operator.h"
+#include "llvm/IR/Verifier.h"
+#include "llvm/Pass.h"
+#include "llvm/PassSupport.h"
+#include "llvm/IR/LegacyPassManager.h"
+#include "llvm/Support/Casting.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Support/ToolOutputFile.h"
+#include "llvm/Transforms/IPO.h"
+
+#include <iostream>
+#include <list>
+#include <memory>
+#include <set>
+#include <sstream>
+#include <vector>
+#include <functional>
+#include <cstdlib>
+
+#define DEBUG_TYPE "spirv"
+
+using namespace llvm;
+using namespace SPIRV;
+using namespace OCLUtil;
+
+namespace llvm {
+  FunctionPass *createPromoteMemoryToRegisterPass();
+}
+
+namespace SPIRV{
+
+cl::opt<bool> SPIRVMemToReg("spirv-mem2reg", cl::init(true),
+    cl::desc("LLVM/SPIR-V translation enable mem2reg"));
+
+
+static void
+foreachKernelArgMD(MDNode *MD, SPIRVFunction *BF,
+    std::function<void(const std::string& Str,
+        SPIRVFunctionParameter *BA)>Func) {
+  for (unsigned I = 1, E = MD->getNumOperands(); I != E; ++I) {
+    SPIRVFunctionParameter *BA = BF->getArgument(I-1);
+    Func(getMDOperandAsString(MD, I), BA);
+  }
+}
+
+/// Information for translating OCL builtin.
+struct OCLBuiltinSPIRVTransInfo {
+  std::string UniqName;
+  /// Postprocessor of operands
+  std::function<void(std::vector<SPIRVWord>&)> PostProc;
+  OCLBuiltinSPIRVTransInfo(){
+    PostProc = [](std::vector<SPIRVWord>&){};
+  }
+};
+
+class LLVMToSPIRVDbgTran {
+public:
+  LLVMToSPIRVDbgTran(Module *TM = nullptr, SPIRVModule *TBM = nullptr)
+  :BM(TBM), M(TM){
+  }
+
+  void setModule(Module *Mod) { M = Mod;}
+  void setSPIRVModule(SPIRVModule *SMod) { BM = SMod;}
+
+  void transDbgInfo(Value *V, SPIRVValue *BV) {
+    if (auto I = dyn_cast<Instruction>(V)) {
+      auto DL = I->getDebugLoc();
+      if (!DL) {
+        //DILocation DIL(*DL.get());
+        auto File = BM->getString((*DL).getFilename().str());
+        // ToDo: SPIR-V rev.31 cannot add debug info for instructions without ids.
+        // This limitation needs to be addressed.
+        if (!BV->hasId())
+          return;
+        BM->addLine(BV, File, (*DL).getLine(), (*DL).getColumn());
+      }
+    } else if (auto F = dyn_cast<Function>(V)) {
+      if (auto DIS = F->getSubprogram()) {
+        auto File = BM->getString(DIS->getFilename().str());
+        BM->addLine(BV, File, DIS->getLine(), 0);
+      }
+    }
+  }
+
+private:
+  SPIRVModule *BM;
+  Module *M;
+};
+
+class LLVMToSPIRV: public ModulePass {
+public:
+  LLVMToSPIRV(SPIRVModule *SMod = nullptr)
+      : ModulePass(ID),
+        M(nullptr),
+        Ctx(nullptr),
+        BM(SMod),
+        ExtSetId(SPIRVID_INVALID),
+        SrcLang(0),
+        SrcLangVer(0),
+        DbgTran(nullptr, SMod){
+  }
+
+  bool runOnModule(Module &Mod) override {
+    M = &Mod;
+    Ctx = &M->getContext();
+    DbgTran.setModule(M);
+    assert(BM && "SPIR-V module not initialized");
+    translate();
+    return true;
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const {
+    AU.addRequired<OCLTypeToSPIRV>();
+  }
+
+  static char ID;
+
+  SPIRVType *transType(Type *T);
+  SPIRVType *transSPIRVOpaqueType(Type *T);
+
+  SPIRVValue *getTranslatedValue(Value *);
+
+  // Translation functions
+  bool transAddressingMode();
+  bool transAlign(Value *V, SPIRVValue *BV);
+  std::vector<SPIRVValue *> transArguments(CallInst *, SPIRVBasicBlock *);
+  std::vector<SPIRVWord> transArguments(CallInst *, SPIRVBasicBlock *,
+      SPIRVEntry *);
+  bool transSourceLanguage();
+  bool transExtension();
+  bool transBuiltinSet();
+  SPIRVValue *transCallInst(CallInst *Call, SPIRVBasicBlock *BB);
+  bool transDecoration(Value *V, SPIRVValue *BV);
+  SPIRVWord transFunctionControlMask(CallInst *);
+  SPIRVWord transFunctionControlMask(Function *);
+  SPIRVFunction *transFunctionDecl(Function *F);
+  bool transGlobalVariables();
+
+  Op transBoolOpCode(SPIRVValue *Opn, Op OC);
+  // Translate LLVM module to SPIR-V module.
+  // Returns true if succeeds.
+  bool translate();
+  bool transExecutionMode();
+  SPIRVValue *transConstant(Value *V);
+  SPIRVValue *transValue(Value *V, SPIRVBasicBlock *BB,
+      bool CreateForward = true);
+  SPIRVValue *transValueWithoutDecoration(Value *V, SPIRVBasicBlock *BB,
+      bool CreateForward = true);
+
+  typedef DenseMap<Type *, SPIRVType *> LLVMToSPIRVTypeMap;
+  typedef DenseMap<Value *, SPIRVValue *> LLVMToSPIRVValueMap;
+private:
+  Module *M;
+  LLVMContext *Ctx;
+  SPIRVModule *BM;
+  LLVMToSPIRVTypeMap TypeMap;
+  LLVMToSPIRVValueMap ValueMap;
+  //ToDo: support multiple builtin sets. Currently assume one builtin set.
+  SPIRVId ExtSetId;
+  SPIRVWord SrcLang;
+  SPIRVWord SrcLangVer;
+  LLVMToSPIRVDbgTran DbgTran;
+
+  SPIRVType *mapType(Type *T, SPIRVType *BT) {
+    TypeMap[T] = BT;
+    SPIRVDBG(dbgs() << "[mapType] " << *T << " => ";
+             spvdbgs() << *BT << '\n');
+    return BT;
+  }
+
+  SPIRVValue *mapValue(Value *V, SPIRVValue *BV) {
+    auto Loc = ValueMap.find(V);
+    if (Loc != ValueMap.end()) {
+      if (Loc->second == BV)
+        return BV;
+      assert (Loc->second->isForward() &&
+          "LLVM Value is mapped to different SPIRV Values");
+      auto Forward = static_cast<SPIRVForward *>(Loc->second);
+      BV->setId(Forward->getId());
+      BM->replaceForward(Forward, BV);
+    }
+    ValueMap[V] = BV;
+    SPIRVDBG(dbgs() << "[mapValue] " << *V << " => ";
+             spvdbgs() << *BV << "\n");
+    return BV;
+  }
+
+  SPIRVType *getSPIRVType(Type *T) {
+      return TypeMap[T];
+  }
+
+  SPIRVValue *getSPIRVValue(Value *V) {
+      return ValueMap[V];
+  }
+
+  SPIRVErrorLog &getErrorLog() {
+    return BM->getErrorLog();
+  }
+
+  llvm::IntegerType* getSizetType();
+  std::vector<SPIRVValue*> transValue(const std::vector<Value *> &Values,
+      SPIRVBasicBlock* BB);
+  std::vector<SPIRVWord> transValue(const std::vector<Value *> &Values,
+      SPIRVBasicBlock* BB, SPIRVEntry *Entry);
+
+  SPIRVInstruction* transBinaryInst(BinaryOperator* B, SPIRVBasicBlock* BB);
+  SPIRVInstruction* transCmpInst(CmpInst* Cmp, SPIRVBasicBlock* BB);
+
+  void dumpUsers(Value *V);
+
+  template<class ExtInstKind>
+  bool oclGetExtInstIndex(const std::string &MangledName,
+      const std::string& DemangledName, SPIRVWord* EntryPoint);
+  void oclGetMutatedArgumentTypesByBuiltin(llvm::FunctionType* FT,
+      std::map<unsigned, Type*>& ChangedType, Function* F);
+
+  bool isBuiltinTransToInst(Function *F);
+  bool isBuiltinTransToExtInst(Function *F,
+      SPIRVExtInstSetKind *BuiltinSet = nullptr,
+      SPIRVWord *EntryPoint = nullptr,
+      SmallVectorImpl<std::string> *Dec = nullptr);
+  bool oclIsKernel(Function *F);
+
+  bool transOCLKernelMetadata();
+
+  SPIRVInstruction *transBuiltinToInst(const std::string& DemangledName,
+      const std::string &MangledName, CallInst* CI, SPIRVBasicBlock* BB);
+  SPIRVInstruction *transBuiltinToInstWithoutDecoration(Op OC,
+      CallInst* CI, SPIRVBasicBlock* BB);
+  void mutateFuncArgType(const std::map<unsigned, Type*>& ChangedType,
+      Function* F);
+
+  SPIRVValue *transSpcvCast(CallInst* CI, SPIRVBasicBlock *BB);
+  SPIRVValue *oclTransSpvcCastSampler(CallInst* CI, SPIRVBasicBlock *BB);
+
+  SPIRV::SPIRVInstruction* transUnaryInst(UnaryInstruction* U,
+      SPIRVBasicBlock* BB);
+
+  /// Add a 32 bit integer constant.
+  /// \return Id of the constant.
+  SPIRVId addInt32(int);
+  void transFunction(Function *I);
+  SPIRV::SPIRVLinkageTypeKind transLinkageType(const GlobalValue* GV);
+};
+
+
+SPIRVValue *
+LLVMToSPIRV::getTranslatedValue(Value *V) {
+  LLVMToSPIRVValueMap::iterator Loc = ValueMap.find(V);
+  if (Loc != ValueMap.end())
+    return Loc->second;
+  return nullptr;
+}
+
+bool
+LLVMToSPIRV::oclIsKernel(Function *F) {
+  if (F->getCallingConv() == CallingConv::SPIR_KERNEL)
+    return true;
+  return false;
+}
+
+bool
+LLVMToSPIRV::isBuiltinTransToInst(Function *F) {
+  std::string DemangledName;
+  if (!oclIsBuiltin(F->getName(), &DemangledName) &&
+      !isDecoratedSPIRVFunc(F, &DemangledName))
+    return false;
+  SPIRVDBG(spvdbgs() << "CallInst: demangled name: " << DemangledName << '\n');
+  return getSPIRVFuncOC(DemangledName) != OpNop;
+}
+
+bool
+LLVMToSPIRV::isBuiltinTransToExtInst(Function *F,
+    SPIRVExtInstSetKind *ExtSet,
+    SPIRVWord *ExtOp,
+    SmallVectorImpl<std::string> *Dec) {
+  std::string OrigName = F->getName();
+  std::string DemangledName;
+  if (!oclIsBuiltin(OrigName, &DemangledName))
+    return false;
+  DEBUG(dbgs() << "[oclIsBuiltinTransToExtInst] CallInst: demangled name: "
+      << DemangledName << '\n');
+  StringRef S = DemangledName;
+  if (!S.startswith(kSPIRVName::Prefix))
+    return false;
+  S = S.drop_front(strlen(kSPIRVName::Prefix));
+  auto Loc = S.find(kSPIRVPostfix::Divider);
+  auto ExtSetName = S.substr(0, Loc);
+  SPIRVExtInstSetKind Set = SPIRVEIS_Count;
+  if (!SPIRVExtSetShortNameMap::rfind(ExtSetName, &Set))
+    return false;
+  assert(Set == BM->getBuiltinSet(ExtSetId) &&
+      "Invalid extended instruction set");
+  assert(Set == SPIRVEIS_OpenCL && "Unsupported extended instruction set");
+
+  auto ExtOpName = S.substr(Loc + 1);
+  auto Splited = ExtOpName.split(kSPIRVPostfix::ExtDivider);
+  OCLExtOpKind EOC;
+  if (!OCLExtOpMap::rfind(Splited.first, &EOC))
+    return false;
+
+  if (ExtSet)
+    *ExtSet = Set;
+  if (ExtOp)
+    *ExtOp = EOC;
+  if (Dec) {
+    SmallVector<StringRef, 2> P;
+    Splited.second.split(P, kSPIRVPostfix::Divider);
+    for (auto &I:P)
+      Dec->push_back(I.str());
+  }
+  return true;
+}
+
+/// Decode SPIR-V type name in the format spirv.{TypeName}._{Postfixes}
+/// where Postfixes are strings separated by underscores.
+/// \return TypeName.
+/// \param Ops contains the integers decoded from postfixes.
+static std::string
+ decodeSPIRVTypeName(StringRef Name,
+    SmallVectorImpl<std::string>& Strs) {
+  SmallVector<StringRef, 4> SubStrs;
+  const char Delim[] = { kSPIRVTypeName::Delimiter, 0 };
+  Name.split(SubStrs, Delim, -1, true);
+  assert(SubStrs.size() >= 2 && "Invalid SPIRV type name");
+  assert(SubStrs[0] == kSPIRVTypeName::Prefix && "Invalid prefix");
+  assert((SubStrs.size() == 2 || !SubStrs[2].empty()) && "Invalid postfix");
+
+  if (SubStrs.size() > 2) {
+    const char PostDelim[] = { kSPIRVTypeName::PostfixDelim, 0 };
+    SmallVector<StringRef, 4> Postfixes;
+    SubStrs[2].split(Postfixes, PostDelim, -1, true);
+    assert(Postfixes.size() > 1 && Postfixes[0].empty() && "Invalid postfix");
+    for (unsigned I = 1, E = Postfixes.size(); I != E; ++I)
+      Strs.push_back(std::string(Postfixes[I]).c_str());
+  }
+  return SubStrs[1].str();
+}
+
+static bool recursiveType(const StructType *ST, const Type *Ty) {
+  SmallPtrSet<const StructType *, 4> Seen;
+
+  std::function<bool(const Type *Ty)> Run = [&](const Type *Ty) {
+    if (!isa<CompositeType>(Ty))
+      return false;
+
+    if (auto *StructTy = dyn_cast<StructType>(Ty)) {
+      if (StructTy == ST)
+        return true;
+
+      if (Seen.count(StructTy))
+        return false;
+
+      Seen.insert(StructTy);
+
+      return find_if(StructTy->subtype_begin(), StructTy->subtype_end(), Run) !=
+             StructTy->subtype_end();
+    }
+
+    if (auto *PtrTy = dyn_cast<PointerType>(Ty))
+      return Run(PtrTy->getPointerElementType());
+
+    if (auto *ArrayTy = dyn_cast<ArrayType>(Ty))
+      return Run(ArrayTy->getArrayElementType());
+
+    return false;
+  };
+
+  return Run(Ty);
+}
+
+SPIRVType *
+LLVMToSPIRV::transType(Type *T) {
+  LLVMToSPIRVTypeMap::iterator Loc = TypeMap.find(T);
+  if (Loc != TypeMap.end())
+    return Loc->second;
+
+  SPIRVDBG(dbgs() << "[transType] " << *T << '\n');
+  if (T->isVoidTy())
+    return mapType(T, BM->addVoidType());
+
+  if (T->isIntegerTy(1))
+    return mapType(T, BM->addBoolType());
+
+  if (T->isIntegerTy())
+    return mapType(T, BM->addIntegerType(T->getIntegerBitWidth()));
+
+  if (T->isFloatingPointTy())
+    return mapType(T, BM->addFloatType(T->getPrimitiveSizeInBits()));
+
+  // A pointer to image or pipe type in LLVM is translated to a SPIRV
+  // sampler or pipe type.
+  if (T->isPointerTy()) {
+    auto ET = T->getPointerElementType();
+    assert(!ET->isFunctionTy() && "Function pointer type is not allowed");
+    auto ST = dyn_cast<StructType>(ET);
+    auto AddrSpc = T->getPointerAddressSpace();
+    if (ST && !ST->isSized()) {
+      Op OpCode;
+      StringRef STName = ST->getName();
+      // Workaround for non-conformant SPIR binary
+      if (STName == "struct._event_t") {
+        STName = kSPR2TypeName::Event;
+        ST->setName(STName);
+      }
+      assert (!STName.startswith(kSPR2TypeName::Pipe) &&
+              "OpenCL type names should be translated to SPIR-V type names");
+      // ToDo: For SPIR1.2/2.0 there may still be load/store or bitcast
+      // instructions using opencl.* type names. We need to handle these
+      // type names until they are all mapped or FE generates SPIR-V type
+      // names.
+      if (STName.find(kSPR2TypeName::Pipe) == 0) {
+        assert(AddrSpc == SPIRAS_Global);
+        SmallVector<StringRef, 4> SubStrs;
+        const char Delims[] = {kSPR2TypeName::Delimiter, 0};
+        STName.split(SubStrs, Delims);
+        std::string Acc = kAccessQualName::ReadOnly;
+        if (SubStrs.size() > 2) {
+          Acc = SubStrs[2];
+        }
+        auto PipeT = BM->addPipeType();
+        PipeT->setPipeAcessQualifier(SPIRSPIRVAccessQualifierMap::map(Acc));
+        return mapType(T, PipeT);
+      } else if (STName.find(kSPR2TypeName::ImagePrefix) == 0) {
+        assert(AddrSpc == SPIRAS_Global);
+        auto SPIRVImageTy = getSPIRVImageTypeFromOCL(M, T);
+        return mapType(T, transSPIRVOpaqueType(SPIRVImageTy));
+      } else if (STName.startswith(kSPIRVTypeName::PrefixAndDelim))
+        return transSPIRVOpaqueType(T);
+      else if (OCLOpaqueTypeOpCodeMap::find(STName, &OpCode)) {
+        if (OpCode == OpTypePipe) {
+          return mapType(T, BM->addPipeType());
+        }
+        return mapType(T, BM->addOpaqueGenericType(OpCode));
+      } else if (isPointerToOpaqueStructType(T)) {
+        return mapType(T, BM->addPointerType(SPIRSPIRVAddrSpaceMap::map(
+          static_cast<SPIRAddressSpace>(AddrSpc)),
+          transType(ET)));
+      }
+    } else  {
+      return mapType(T, BM->addPointerType(SPIRSPIRVAddrSpaceMap::map(
+        static_cast<SPIRAddressSpace>(AddrSpc)),
+        transType(ET)));
+    }
+  }
+
+  if (T->isVectorTy())
+    return mapType(T, BM->addVectorType(transType(T->getVectorElementType()),
+        T->getVectorNumElements()));
+
+  if (T->isArrayTy())
+    return mapType(T, BM->addArrayType(transType(T->getArrayElementType()),
+        static_cast<SPIRVConstant*>(transValue(ConstantInt::get(getSizetType(),
+            T->getArrayNumElements(), false), nullptr))));
+
+  if (T->isStructTy() && !T->isSized()) {
+    auto ST = dyn_cast<StructType>(T);
+    assert(!ST->getName().startswith(kSPR2TypeName::Pipe));
+    assert(!ST->getName().startswith(kSPR2TypeName::ImagePrefix));
+    return mapType(T, BM->addOpaqueType(T->getStructName()));
+  }
+
+  if (auto ST = dyn_cast<StructType>(T)) {
+    assert(ST->isSized());
+
+    std::string Name;
+    if (ST->hasName())
+      Name = ST->getName();
+
+    if(Name == getSPIRVTypeName(kSPIRVTypeName::ConstantSampler))
+      return transType(getSamplerType(M));
+
+    auto *Struct = BM->openStructType(T->getStructNumElements(), Name);
+    mapType(T, Struct);
+
+    SmallVector<unsigned, 4> ForwardRefs;
+
+    for (unsigned I = 0, E = T->getStructNumElements(); I != E; ++I) {
+      auto *ElemTy = ST->getElementType(I);
+      if (isa<CompositeType>(ElemTy) && recursiveType(ST, ElemTy))
+        ForwardRefs.push_back(I);
+      else
+        Struct->setMemberType(I, transType(ST->getElementType(I)));
+    }
+
+    BM->closeStructType(Struct, ST->isPacked());
+
+    for (auto I : ForwardRefs)
+      Struct->setMemberType(I, transType(ST->getElementType(I)));
+
+    return Struct;
+  }
+
+  if (FunctionType *FT = dyn_cast<FunctionType>(T)) {
+    SPIRVType *RT = transType(FT->getReturnType());
+    std::vector<SPIRVType *> PT;
+    for (FunctionType::param_iterator I = FT->param_begin(),
+        E = FT->param_end(); I != E; ++I)
+      PT.push_back(transType(*I));
+    return mapType(T, BM->addFunctionType(RT, PT));
+  }
+
+  llvm_unreachable("Not implemented!");
+  return 0;
+}
+
+SPIRVType *
+LLVMToSPIRV::transSPIRVOpaqueType(Type *T) {
+  auto ET = T->getPointerElementType();
+  auto ST = cast<StructType>(ET);
+  auto AddrSpc = T->getPointerAddressSpace();
+  auto STName = ST->getStructName();
+  assert (STName.startswith(kSPIRVTypeName::PrefixAndDelim) &&
+    "Invalid SPIR-V opaque type name");
+  SmallVector<std::string, 8> Postfixes;
+  auto TN = decodeSPIRVTypeName(STName, Postfixes);
+  if (TN == kSPIRVTypeName::Pipe) {
+    assert(AddrSpc == SPIRAS_Global);
+    assert(Postfixes.size() == 1 && "Invalid pipe type ops");
+    auto PipeT = BM->addPipeType();
+    PipeT->setPipeAcessQualifier(static_cast<spv::AccessQualifier>(
+      atoi(Postfixes[0].c_str())));
+    return mapType(T, PipeT);
+  } else if (TN == kSPIRVTypeName::Image) {
+    assert(AddrSpc == SPIRAS_Global);
+    // The sampled type needs to be translated through LLVM type to guarantee
+    // uniqueness.
+    auto SampledT = transType(getLLVMTypeForSPIRVImageSampledTypePostfix(
+      Postfixes[0], *Ctx));
+    SmallVector<int, 7> Ops;
+    for (unsigned I = 1; I < 8; ++I)
+      Ops.push_back(atoi(Postfixes[I].c_str()));
+    SPIRVTypeImageDescriptor Desc(static_cast<SPIRVImageDimKind>(Ops[0]),
+        Ops[1], Ops[2], Ops[3], Ops[4], Ops[5]);
+    return mapType(T, BM->addImageType(SampledT, Desc,
+                   static_cast<spv::AccessQualifier>(Ops[6])));
+  } else if (TN == kSPIRVTypeName::SampledImg) {
+    return mapType(T, BM->addSampledImageType(
+        static_cast<SPIRVTypeImage *>(
+            transType(getSPIRVTypeByChangeBaseTypeName(M,
+                T, kSPIRVTypeName::SampledImg,
+                kSPIRVTypeName::Image)))));
+  } else if(TN == kSPIRVTypeName::Sampler)
+    return mapType(T, BM->addSamplerType());
+  else if (TN == kSPIRVTypeName::DeviceEvent)
+    return mapType(T, BM->addDeviceEventType());
+  else
+    return mapType(T, BM->addOpaqueGenericType(
+      SPIRVOpaqueTypeOpCodeMap::map(TN)));
+}
+
+SPIRVFunction *
+LLVMToSPIRV::transFunctionDecl(Function *F) {
+  if (auto BF= getTranslatedValue(F))
+    return static_cast<SPIRVFunction *>(BF);
+
+  SPIRVTypeFunction *BFT = static_cast<SPIRVTypeFunction *>(transType(
+      getAnalysis<OCLTypeToSPIRV>().getAdaptedType(F)));
+  SPIRVFunction *BF = static_cast<SPIRVFunction *>(mapValue(F,
+      BM->addFunction(BFT)));
+  BF->setFunctionControlMask(transFunctionControlMask(F));
+  if (F->hasName())
+    BM->setName(BF, F->getName());
+  if (oclIsKernel(F))
+    BM->addEntryPoint(ExecutionModelKernel, BF->getId());
+  else if (F->getLinkage() != GlobalValue::InternalLinkage)
+    BF->setLinkageType(transLinkageType(F));
+  auto Attrs = F->getAttributes();
+  for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
+      ++I) {
+    auto ArgNo = I->getArgNo();
+    SPIRVFunctionParameter *BA = BF->getArgument(ArgNo);
+    if (I->hasName())
+      BM->setName(BA, I->getName());
+    if (I->hasByValAttr())
+      BA->addAttr(FunctionParameterAttributeByVal);
+    if (I->hasNoAliasAttr())
+      BA->addAttr(FunctionParameterAttributeNoAlias);
+    if (I->hasNoCaptureAttr())
+      BA->addAttr(FunctionParameterAttributeNoCapture);
+    if (I->hasStructRetAttr())
+      BA->addAttr(FunctionParameterAttributeSret);
+    if (Attrs.hasAttribute(ArgNo + 1, Attribute::ZExt))
+      BA->addAttr(FunctionParameterAttributeZext);
+    if (Attrs.hasAttribute(ArgNo + 1, Attribute::SExt))
+      BA->addAttr(FunctionParameterAttributeSext);
+  }
+  if (Attrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt))
+    BF->addDecorate(DecorationFuncParamAttr, FunctionParameterAttributeZext);
+  if (Attrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
+    BF->addDecorate(DecorationFuncParamAttr, FunctionParameterAttributeSext);
+  DbgTran.transDbgInfo(F, BF);
+  SPIRVDBG(dbgs() << "[transFunction] " << *F << " => ";
+    spvdbgs() << *BF << '\n';)
+  return BF;
+}
+
+#define _SPIRV_OPL(x) OpLogical##x
+
+#define _SPIRV_OPB(x) OpBitwise##x
+
+SPIRVValue *
+LLVMToSPIRV::transConstant(Value *V) {
+  if (auto CPNull = dyn_cast<ConstantPointerNull>(V))
+    return BM->addNullConstant(bcast<SPIRVTypePointer>(transType(
+        CPNull->getType())));
+
+  if (auto CAZero = dyn_cast<ConstantAggregateZero>(V)) {
+    Type *AggType = CAZero->getType();
+    if (const StructType* ST = dyn_cast<StructType>(AggType)) {
+      if (ST->getName() == getSPIRVTypeName(kSPIRVTypeName::ConstantSampler))
+        return BM->addSamplerConstant(transType(AggType), 0,0,0);
+    } else
+      return BM->addNullConstant(transType(AggType));
+  }
+
+  if (auto ConstI = dyn_cast<ConstantInt>(V))
+    return BM->addConstant(transType(V->getType()), ConstI->getZExtValue());
+
+  if (auto ConstFP = dyn_cast<ConstantFP>(V)) {
+    auto BT = static_cast<SPIRVType *>(transType(V->getType()));
+    return BM->addConstant(BT,
+        ConstFP->getValueAPF().bitcastToAPInt().getZExtValue());
+  }
+
+  if (auto ConstDA = dyn_cast<ConstantDataArray>(V)) {
+    std::vector<SPIRVValue *> BV;
+    for (unsigned I = 0, E = ConstDA->getNumElements(); I != E; ++I)
+      BV.push_back(transValue(ConstDA->getElementAsConstant(I), nullptr));
+    return BM->addCompositeConstant(transType(V->getType()), BV);
+  }
+
+  if (auto ConstA = dyn_cast<ConstantArray>(V)) {
+    std::vector<SPIRVValue *> BV;
+    for (auto I = ConstA->op_begin(), E = ConstA->op_end(); I != E; ++I)
+      BV.push_back(transValue(*I, nullptr));
+    return BM->addCompositeConstant(transType(V->getType()), BV);
+  }
+
+  if (auto ConstDV = dyn_cast<ConstantDataVector>(V)) {
+    std::vector<SPIRVValue *> BV;
+    for (unsigned I = 0, E = ConstDV->getNumElements(); I != E; ++I)
+      BV.push_back(transValue(ConstDV->getElementAsConstant(I), nullptr));
+    return BM->addCompositeConstant(transType(V->getType()), BV);
+  }
+
+  if (auto ConstV = dyn_cast<ConstantVector>(V)) {
+    std::vector<SPIRVValue *> BV;
+    for (auto I = ConstV->op_begin(), E = ConstV->op_end(); I != E; ++I)
+      BV.push_back(transValue(*I, nullptr));
+    return BM->addCompositeConstant(transType(V->getType()), BV);
+  }
+
+  if (auto ConstV = dyn_cast<ConstantStruct>(V)) {
+    if (ConstV->getType()->getName() ==
+        getSPIRVTypeName(kSPIRVTypeName::ConstantSampler)) {
+      assert(ConstV->getNumOperands() == 3);
+      SPIRVWord
+        AddrMode   = ConstV->getOperand(0)->getUniqueInteger().getZExtValue(),
+        Normalized = ConstV->getOperand(1)->getUniqueInteger().getZExtValue(),
+        FilterMode = ConstV->getOperand(2)->getUniqueInteger().getZExtValue();
+      assert(AddrMode < 5 && "Invalid addressing mode");
+      assert(Normalized < 2 && "Invalid value of normalized coords");
+      assert(FilterMode < 2 && "Invalid filter mode");
+      SPIRVType* SamplerTy = transType(ConstV->getType());
+      return BM->addSamplerConstant(SamplerTy,
+                                    AddrMode, Normalized, FilterMode);
+    }
+    std::vector<SPIRVValue *> BV;
+    for (auto I = ConstV->op_begin(), E = ConstV->op_end(); I != E; ++I)
+      BV.push_back(transValue(*I, nullptr));
+    return BM->addCompositeConstant(transType(V->getType()), BV);
+  }
+
+  if (auto ConstUE = dyn_cast<ConstantExpr>(V)) {
+    auto Inst = ConstUE->getAsInstruction();
+    SPIRVDBG(dbgs() << "ConstantExpr: " << *ConstUE << '\n';
+      dbgs() << "Instruction: " << *Inst << '\n';)
+    auto BI = transValue(Inst, nullptr, false);
+    Inst->dropAllReferences();
+    return BI;
+  }
+
+  if (isa<UndefValue>(V)) {
+    return BM->addUndef(transType(V->getType()));
+  }
+
+  return nullptr;
+}
+
+SPIRVValue *
+LLVMToSPIRV::transValue(Value *V, SPIRVBasicBlock *BB, bool CreateForward) {
+  LLVMToSPIRVValueMap::iterator Loc = ValueMap.find(V);
+  if (Loc != ValueMap.end() && (!Loc->second->isForward() || CreateForward))
+    return Loc->second;
+
+  SPIRVDBG(dbgs() << "[transValue] " << *V << '\n');
+  assert ((!isa<Instruction>(V) || isa<GetElementPtrInst>(V) ||
+      isa<CastInst>(V) || BB) &&
+      "Invalid SPIRV BB");
+
+  auto BV = transValueWithoutDecoration(V, BB, CreateForward);
+  std::string name = V->getName();
+  if (!name.empty()) // Don't erase the name, which BM might already have
+    BM->setName(BV, name);
+  if(!transDecoration(V, BV))
+    return nullptr;
+  return BV;
+}
+
+SPIRVInstruction*
+LLVMToSPIRV::transBinaryInst(BinaryOperator* B, SPIRVBasicBlock* BB) {
+  unsigned LLVMOC = B->getOpcode();
+  auto Op0 = transValue(B->getOperand(0), BB);
+  SPIRVInstruction* BI = BM->addBinaryInst(
+      transBoolOpCode(Op0, OpCodeMap::map(LLVMOC)),
+      transType(B->getType()), Op0, transValue(B->getOperand(1), BB), BB);
+  return BI;
+}
+
+SPIRVInstruction*
+LLVMToSPIRV::transCmpInst(CmpInst* Cmp, SPIRVBasicBlock* BB) {
+  auto Op0 = transValue(Cmp->getOperand(0), BB);
+  SPIRVInstruction* BI = BM->addCmpInst(
+      transBoolOpCode(Op0, CmpMap::map(Cmp->getPredicate())),
+      transType(Cmp->getType()), Op0,
+      transValue(Cmp->getOperand(1), BB), BB);
+  return BI;
+}
+
+SPIRV::SPIRVInstruction *LLVMToSPIRV::transUnaryInst(UnaryInstruction *U,
+                                                  SPIRVBasicBlock *BB) {
+  Op BOC = OpNop;
+  if (auto Cast = dyn_cast<AddrSpaceCastInst>(U)) {
+    if (Cast->getDestTy()->getPointerAddressSpace() == SPIRAS_Generic) {
+      assert(Cast->getSrcTy()->getPointerAddressSpace() != SPIRAS_Constant &&
+             "Casts from constant address space to generic are illegal");
+      BOC = OpPtrCastToGeneric;
+    } else {
+      assert(Cast->getDestTy()->getPointerAddressSpace() != SPIRAS_Constant &&
+             "Casts from generic address space to constant are illegal");
+      assert(Cast->getSrcTy()->getPointerAddressSpace() == SPIRAS_Generic);
+      BOC = OpGenericCastToPtr;
+    }
+  } else {
+    auto OpCode = U->getOpcode();
+    BOC = OpCodeMap::map(OpCode);
+  }
+
+  auto Op = transValue(U->getOperand(0), BB);
+  return BM->addUnaryInst(transBoolOpCode(Op, BOC),
+      transType(U->getType()), Op, BB);
+}
+
+/// An instruction may use an instruction from another BB which has not been
+/// translated. SPIRVForward should be created as place holder for these
+/// instructions and replaced later by the real instructions.
+/// Use CreateForward = true to indicate such situation.
+SPIRVValue *
+LLVMToSPIRV::transValueWithoutDecoration(Value *V, SPIRVBasicBlock *BB,
+    bool CreateForward) {
+  if (auto LBB = dyn_cast<BasicBlock>(V)) {
+    auto BF = static_cast<SPIRVFunction *>(getTranslatedValue(LBB->getParent()));
+    assert (BF && "Function not translated");
+    BB = static_cast<SPIRVBasicBlock *>(mapValue(V, BM->addBasicBlock(BF)));
+    BM->setName(BB, LBB->getName());
+    return BB;
+  }
+
+  if (auto F = dyn_cast<Function>(V))
+    return transFunctionDecl(F);
+
+  if (auto GV = dyn_cast<GlobalVariable>(V)) {
+    llvm::PointerType * Ty = GV->getType();
+    // Though variables with common linkage type are initialized by 0,
+    // they can be represented in SPIR-V as uninitialized variables with
+    // 'Export' linkage type, just as tentative definitions look in C
+    llvm::Value *Init = GV->hasInitializer() && !GV->hasCommonLinkage() ?
+      GV->getInitializer() : nullptr;
+    StructType *ST = Init ? dyn_cast<StructType>(Init->getType()) : nullptr;
+    if (ST && ST->hasName() &&
+        ST->getName() == getSPIRVTypeName(kSPIRVTypeName::ConstantSampler)) {
+      auto BV = transConstant(Init);
+      assert(BV);
+      return mapValue(V, BV);
+    } else if (ConstantExpr *ConstUE = dyn_cast_or_null<ConstantExpr>(Init)) {
+      Instruction * Inst = ConstUE->getAsInstruction();
+      if (isSamplerInitializer(Inst)) {
+        Init = Inst->getOperand(0);
+        Ty = static_cast<PointerType*>(Init->getType());
+      }
+      Inst->dropAllReferences();
+    }
+    auto BVar = static_cast<SPIRVVariable *>(BM->addVariable(
+      transType(Ty), GV->isConstant(),
+      transLinkageType(GV),
+      Init ? transValue(Init, nullptr) : nullptr,
+      GV->getName(),
+      SPIRSPIRVAddrSpaceMap::map(
+        static_cast<SPIRAddressSpace>(Ty->getAddressSpace())),
+      nullptr
+      ));
+    mapValue(V, BVar);
+    spv::BuiltIn Builtin = spv::BuiltInPosition;
+    if (!GV->hasName() || !getSPIRVBuiltin(GV->getName().str(), Builtin))
+      return BVar;
+    BVar->setBuiltin(Builtin);
+    return BVar;
+  }
+
+  if (isa<Constant>(V)) {
+    auto BV = transConstant(V);
+    assert(BV);
+    return mapValue(V, BV);
+  }
+
+  if (auto Arg = dyn_cast<Argument>(V)) {
+    unsigned ArgNo = Arg->getArgNo();
+    SPIRVFunction *BF = BB->getParent();
+    //assert(BF->existArgument(ArgNo));
+    return mapValue(V, BF->getArgument(ArgNo));
+  }
+
+  if (CreateForward)
+    return mapValue(V, BM->addForward(transType(V->getType())));
+
+  if (StoreInst *ST = dyn_cast<StoreInst>(V)) {
+    std::vector<SPIRVWord> MemoryAccess;
+    if (ST->isVolatile())
+      MemoryAccess.push_back(MemoryAccessVolatileMask);
+    MemoryAccess.push_back(MemoryAccessAlignedMask);
+    MemoryAccess.push_back(ST->getAlignment());
+    return mapValue(V, BM->addStoreInst(
+        transValue(ST->getPointerOperand(), BB),
+        transValue(ST->getValueOperand(), BB),
+        MemoryAccess, BB));
+  }
+
+  if (LoadInst *LD = dyn_cast<LoadInst>(V)) {
+    std::vector<SPIRVWord> MemoryAccess;
+    if (LD->isVolatile())
+      MemoryAccess.push_back(MemoryAccessVolatileMask);
+    MemoryAccess.push_back(MemoryAccessAlignedMask);
+    MemoryAccess.push_back(LD->getAlignment());
+    return mapValue(V, BM->addLoadInst(
+        transValue(LD->getPointerOperand(), BB),
+        MemoryAccess, BB));
+  }
+
+  if (BinaryOperator *B = dyn_cast<BinaryOperator>(V)) {
+    SPIRVInstruction* BI = transBinaryInst(B, BB);
+    return mapValue(V, BI);
+  }
+
+  if (auto RI = dyn_cast<ReturnInst>(V)) {
+    if (auto RV = RI->getReturnValue())
+      return mapValue(V, BM->addReturnValueInst(
+          transValue(RV, BB), BB));
+    return mapValue(V, BM->addReturnInst(BB));
+  }
+
+  if (CmpInst *Cmp = dyn_cast<CmpInst>(V)) {
+    SPIRVInstruction* BI = transCmpInst(Cmp, BB);
+    return mapValue(V, BI);
+  }
+
+  if (SelectInst *Sel = dyn_cast<SelectInst>(V))
+    return mapValue(V, BM->addSelectInst(
+        transValue(Sel->getCondition(), BB),
+        transValue(Sel->getTrueValue(), BB),
+        transValue(Sel->getFalseValue(), BB),BB));
+
+  if (AllocaInst *Alc = dyn_cast<AllocaInst>(V))
+    return mapValue(V, BM->addVariable(
+      transType(Alc->getType()), false,
+      SPIRVLinkageTypeKind::LinkageTypeInternal,
+      nullptr, Alc->getName(),
+      StorageClassFunction, BB));
+
+  if (auto *Switch = dyn_cast<SwitchInst>(V)) {
+    std::vector<std::pair<SPIRVWord, SPIRVBasicBlock *>> Pairs;
+    for (auto I = Switch->case_begin(), E = Switch->case_end(); I != E; ++I)
+      Pairs.push_back(std::make_pair(I.getCaseValue()->getZExtValue(),
+          static_cast<SPIRVBasicBlock*>(transValue(I.getCaseSuccessor(),
+              nullptr))));
+    return mapValue(V, BM->addSwitchInst(
+        transValue(Switch->getCondition(), BB),
+        static_cast<SPIRVBasicBlock*>(transValue(Switch->getDefaultDest(),
+            nullptr)), Pairs, BB));
+  }
+
+  if (auto Branch = dyn_cast<BranchInst>(V)) {
+    if (Branch->isUnconditional())
+      return mapValue(V, BM->addBranchInst(
+          static_cast<SPIRVLabel*>(transValue(Branch->getSuccessor(0), BB)),
+          BB));
+    return mapValue(V, BM->addBranchConditionalInst(
+        transValue(Branch->getCondition(), BB),
+        static_cast<SPIRVLabel*>(transValue(Branch->getSuccessor(0), BB)),
+        static_cast<SPIRVLabel*>(transValue(Branch->getSuccessor(1), BB)),
+        BB));
+  }
+
+  if (auto Phi = dyn_cast<PHINode>(V)) {
+    std::vector<SPIRVValue *> IncomingPairs;
+    for (size_t I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) {
+      IncomingPairs.push_back(transValue(Phi->getIncomingValue(I), BB));
+      IncomingPairs.push_back(transValue(Phi->getIncomingBlock(I), nullptr));
+    }
+    return mapValue(V, BM->addPhiInst(transType(Phi->getType()), IncomingPairs,
+        BB));
+  }
+
+  if (UnaryInstruction *U = dyn_cast<UnaryInstruction>(V)) {
+    if(isSamplerInitializer(U))
+      return mapValue(V, transValue(U->getOperand(0), BB));
+    return mapValue(V, transUnaryInst(U, BB));
+  }
+
+  if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
+    std::vector<SPIRVValue *> Indices;
+    for (unsigned i = 0, e = GEP->getNumIndices(); i != e; ++i)
+      Indices.push_back(transValue(GEP->getOperand(i+1), BB));
+    return mapValue(V, BM->addPtrAccessChainInst(
+        transType(GEP->getType()),
+        transValue(GEP->getPointerOperand(), BB),
+        Indices, BB, GEP->isInBounds()));
+  }
+
+  if (auto Ext = dyn_cast<ExtractElementInst>(V)) {
+    auto Index = Ext->getIndexOperand();
+    if (auto Const = dyn_cast<ConstantInt>(Index))
+      return mapValue(V, BM->addCompositeExtractInst(
+        transType(Ext->getType()),
+        transValue(Ext->getVectorOperand(), BB),
+        std::vector<SPIRVWord>(1, Const->getZExtValue()),
+        BB));
+    else
+      return mapValue(V, BM->addVectorExtractDynamicInst(
+          transValue(Ext->getVectorOperand(), BB),
+          transValue(Index, BB),
+          BB));
+  }
+
+  if (auto Ins = dyn_cast<InsertElementInst>(V)) {
+    auto Index = Ins->getOperand(2);
+    if (auto Const = dyn_cast<ConstantInt>(Index))
+      return mapValue(V, BM->addCompositeInsertInst(
+      transValue(Ins->getOperand(1), BB),
+      transValue(Ins->getOperand(0), BB),
+      std::vector<SPIRVWord>(1, Const->getZExtValue()),
+      BB));
+    else
+      return mapValue(V, BM->addVectorInsertDynamicInst(
+      transValue(Ins->getOperand(0), BB),
+      transValue(Ins->getOperand(1), BB),
+      transValue(Index, BB),
+      BB));
+  }
+
+  if (auto SF = dyn_cast<ShuffleVectorInst>(V)) {
+    std::vector<SPIRVWord> Comp;
+    for (auto &I:SF->getShuffleMask())
+      Comp.push_back(I);
+    return mapValue(V, BM->addVectorShuffleInst(
+        transType(SF->getType()),
+        transValue(SF->getOperand(0), BB),
+        transValue(SF->getOperand(1), BB),
+        Comp,
+        BB));
+  }
+
+  if (CallInst *CI = dyn_cast<CallInst>(V))
+    return mapValue(V, transCallInst(CI, BB));
+
+  llvm_unreachable("Not implemented");
+  return nullptr;
+}
+
+bool
+LLVMToSPIRV::transDecoration(Value *V, SPIRVValue *BV) {
+  if (!transAlign(V, BV))
+    return false;
+  if ((isa<AtomicCmpXchgInst>(V) &&
+      cast<AtomicCmpXchgInst>(V)->isVolatile()) ||
+      (isa<AtomicRMWInst>(V) && cast<AtomicRMWInst>(V)->isVolatile()))
+    BV->setVolatile(true);
+  DbgTran.transDbgInfo(V, BV);
+  return true;
+}
+
+bool
+LLVMToSPIRV::transAlign(Value *V, SPIRVValue *BV) {
+  if (auto AL = dyn_cast<AllocaInst>(V)) {
+    BM->setAlignment(BV, AL->getAlignment());
+    return true;
+  }
+  if (auto GV = dyn_cast<GlobalVariable>(V)) {
+    BM->setAlignment(BV, GV->getAlignment());
+    return true;
+  }
+  return true;
+}
+
+/// Do this after source language is set.
+bool
+LLVMToSPIRV::transBuiltinSet() {
+  SPIRVWord Ver = 0;
+  SourceLanguage Kind = BM->getSourceLanguage(&Ver);
+  assert((Kind == SourceLanguageOpenCL_C ||
+      Kind == SourceLanguageOpenCL_CPP ) && "not supported");
+  std::stringstream SS;
+  SS << "OpenCL.std";
+  return BM->importBuiltinSet(SS.str(), &ExtSetId);
+}
+
+/// Transform sampler* spcv.cast(i32 arg)
+/// Only two cases are possible:
+///   arg = ConstantInt x -> SPIRVConstantSampler
+///   arg = i32 argument -> transValue(arg)
+///   arg = load from sampler -> look through load
+SPIRVValue *
+LLVMToSPIRV::oclTransSpvcCastSampler(CallInst* CI, SPIRVBasicBlock *BB) {
+  llvm::Function* F = CI->getCalledFunction();
+  auto FT = F->getFunctionType();
+  auto RT = FT->getReturnType();
+  assert(FT->getNumParams() == 1);
+  assert(isSPIRVType(RT, kSPIRVTypeName::Sampler) &&
+    FT->getParamType(0)->isIntegerTy() && "Invalid sampler type");
+  auto Arg = CI->getArgOperand(0);
+
+  auto GetSamplerConstant = [&](uint64_t SamplerValue) {
+    auto AddrMode = (SamplerValue & 0xE) >> 1;
+    auto Param = SamplerValue & 0x1;
+    auto Filter = ((SamplerValue & 0x30) >> 4) - 1;
+    auto BV = BM->addSamplerConstant(transType(RT), AddrMode, Param, Filter);
+    return BV;
+  };
+
+  if (auto Const = dyn_cast<ConstantInt>(Arg)) {
+    // Sampler is declared as a kernel scope constant
+    return GetSamplerConstant(Const->getZExtValue());
+  } else if (auto Load = dyn_cast<LoadInst>(Arg)) {
+    // If value of the sampler is loaded from a global constant, use its
+    // initializer for initialization of the sampler.
+    auto Op = Load->getPointerOperand();
+    assert(isa<GlobalVariable>(Op) && "Unknown sampler pattern!");
+    auto GV = cast<GlobalVariable>(Op);
+    assert(GV->isConstant() ||
+      GV->getType()->getPointerAddressSpace() == SPIRAS_Constant);
+    auto Initializer = GV->getInitializer();
+    assert(isa<ConstantInt>(Initializer) && "sampler not constant int?");
+    return GetSamplerConstant(cast<ConstantInt>(Initializer)->getZExtValue());
+  }
+  // Sampler is a function argument
+  auto BV = transValue(Arg, BB);
+  assert(BV && BV->getType() == transType(RT));
+  return BV;
+}
+
+SPIRVValue *
+LLVMToSPIRV::transSpcvCast(CallInst* CI, SPIRVBasicBlock *BB) {
+  return oclTransSpvcCastSampler(CI, BB);
+}
+
+SPIRVValue *
+LLVMToSPIRV::transCallInst(CallInst *CI, SPIRVBasicBlock *BB) {
+  SPIRVExtInstSetKind ExtSetKind = SPIRVEIS_Count;
+  SPIRVWord ExtOp = SPIRVWORD_MAX;
+  llvm::Function* F = CI->getCalledFunction();
+  auto MangledName = F->getName();
+  std::string DemangledName;
+
+  if (MangledName.startswith(SPCV_CAST))
+    return transSpcvCast(CI, BB);
+
+  if (MangledName.startswith("llvm.memcpy")) {
+    std::vector<SPIRVWord> MemoryAccess;
+
+    if (isa<ConstantInt>(CI->getOperand(4)) &&
+      dyn_cast<ConstantInt>(CI->getOperand(4))
+      ->getZExtValue() == 1)
+      MemoryAccess.push_back(MemoryAccessVolatileMask);
+    if (isa<ConstantInt>(CI->getOperand(3))) {
+        MemoryAccess.push_back(MemoryAccessAlignedMask);
+        MemoryAccess.push_back(dyn_cast<ConstantInt>(CI->getOperand(3))
+          ->getZExtValue());
+    }
+
+    return BM->addCopyMemorySizedInst(
+      transValue(CI->getOperand(0), BB),
+      transValue(CI->getOperand(1), BB),
+      transValue(CI->getOperand(2), BB),
+      MemoryAccess,
+      BB);
+  }
+
+  if (oclIsBuiltin(MangledName, &DemangledName) ||
+      isDecoratedSPIRVFunc(F, &DemangledName))
+    if (auto BV = transBuiltinToInst(DemangledName, MangledName, CI, BB))
+      return BV;
+
+  SmallVector<std::string, 2> Dec;
+  if (isBuiltinTransToExtInst(CI->getCalledFunction(), &ExtSetKind,
+      &ExtOp, &Dec))
+    return addDecorations(BM->addExtInst(
+      transType(CI->getType()),
+      ExtSetId,
+      ExtOp,
+      transArguments(CI, BB, SPIRVEntry::create_unique(ExtSetKind, ExtOp).get()),
+      BB), Dec);
+
+  return BM->addCallInst(
+    transFunctionDecl(CI->getCalledFunction()),
+    transArguments(CI, BB, SPIRVEntry::create_unique(OpFunctionCall).get()),
+    BB);
+}
+
+bool
+LLVMToSPIRV::transAddressingMode() {
+  Triple TargetTriple(M->getTargetTriple());
+  Triple::ArchType Arch = TargetTriple.getArch();
+
+  SPIRVCKRT(Arch == Triple::spir || Arch == Triple::spir64,
+      InvalidTargetTriple,
+      "Actual target triple is " + M->getTargetTriple());
+
+  if (Arch == Triple::spir)
+    BM->setAddressingModel(AddressingModelPhysical32);
+  else
+    BM->setAddressingModel(AddressingModelPhysical64);
+  // Physical addressing model requires Addresses capability
+  BM->addCapability(CapabilityAddresses);
+  return true;
+}
+std::vector<SPIRVValue*>
+LLVMToSPIRV::transValue(const std::vector<Value *> &Args, SPIRVBasicBlock* BB) {
+  std::vector<SPIRVValue*> BArgs;
+  for (auto &I: Args)
+    BArgs.push_back(transValue(I, BB));
+  return BArgs;
+}
+
+std::vector<SPIRVValue*>
+LLVMToSPIRV::transArguments(CallInst *CI, SPIRVBasicBlock *BB) {
+  return transValue(getArguments(CI), BB);
+}
+
+std::vector<SPIRVWord>
+LLVMToSPIRV::transValue(const std::vector<Value *> &Args, SPIRVBasicBlock* BB,
+    SPIRVEntry *Entry) {
+  std::vector<SPIRVWord> Operands;
+  for (size_t I = 0, E = Args.size(); I != E; ++I) {
+    Operands.push_back(Entry->isOperandLiteral(I) ?
+        cast<ConstantInt>(Args[I])->getZExtValue() :
+        transValue(Args[I], BB)->getId());
+  }
+  return Operands;
+}
+
+std::vector<SPIRVWord>
+LLVMToSPIRV::transArguments(CallInst *CI, SPIRVBasicBlock *BB, SPIRVEntry *Entry) {
+  return transValue(getArguments(CI), BB, Entry);
+}
+
+SPIRVWord
+LLVMToSPIRV::transFunctionControlMask(CallInst *CI) {
+  SPIRVWord FCM = 0;
+  SPIRSPIRVFuncCtlMaskMap::foreach([&](Attribute::AttrKind Attr,
+      SPIRVFunctionControlMaskKind Mask){
+    if (CI->hasFnAttr(Attr))
+      FCM |= Mask;
+  });
+  return FCM;
+}
+
+SPIRVWord
+LLVMToSPIRV::transFunctionControlMask(Function *F) {
+  SPIRVWord FCM = 0;
+  SPIRSPIRVFuncCtlMaskMap::foreach([&](Attribute::AttrKind Attr,
+      SPIRVFunctionControlMaskKind Mask){
+    if (F->hasFnAttribute(Attr))
+      FCM |= Mask;
+  });
+  return FCM;
+}
+
+bool
+LLVMToSPIRV::transGlobalVariables() {
+  for (auto I = M->global_begin(),
+            E = M->global_end(); I != E; ++I) {
+    if (!transValue(&*I, nullptr))
+      return false;
+  }
+  return true;
+}
+
+void
+LLVMToSPIRV::mutateFuncArgType(const std::map<unsigned, Type*>& ChangedType,
+    Function* F) {
+  for (auto &I : ChangedType) {
+    for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE; ++UI) {
+      auto Call = dyn_cast<CallInst>(*UI);
+      if (!Call)
+        continue;
+      auto Arg = Call->getArgOperand(I.first);
+      auto OrigTy = Arg->getType();
+      if (OrigTy == I.second)
+        continue;
+      SPIRVDBG(dbgs() << "[mutate arg type] " << *Call << ", " << *Arg << '\n');
+      auto CastF = M->getOrInsertFunction(SPCV_CAST, I.second, OrigTy, nullptr);
+      std::vector<Value *> Args;
+      Args.push_back(Arg);
+      auto Cast = CallInst::Create(CastF, Args, "", Call);
+      Call->replaceUsesOfWith(Arg, Cast);
+      SPIRVDBG(dbgs() << "[mutate arg type] -> " << *Cast << '\n');
+    }
+  }
+}
+
+void
+LLVMToSPIRV::transFunction(Function *I) {
+  transFunctionDecl(I);
+  // Creating all basic blocks before creating any instruction.
+  for (Function::iterator FI = I->begin(), FE = I->end(); FI != FE; ++FI) {
+    transValue(&*FI, nullptr);
+  }
+  for (Function::iterator FI = I->begin(), FE = I->end(); FI != FE; ++FI) {
+    SPIRVBasicBlock* BB = static_cast<SPIRVBasicBlock*>(transValue(&*FI, nullptr));
+    for (BasicBlock::iterator BI = FI->begin(), BE = FI->end(); BI != BE;
+        ++BI) {
+      transValue(&*BI, BB, false);
+    }
+  }
+}
+
+bool
+LLVMToSPIRV::translate() {
+  BM->setGeneratorVer(kTranslatorVer);
+
+  if (!transSourceLanguage())
+    return false;
+  if (!transExtension())
+    return false;
+  if (!transBuiltinSet())
+    return false;
+  if (!transAddressingMode())
+    return false;
+  if (!transGlobalVariables())
+    return false;
+
+  for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
+    Function *F = &*I;
+    auto FT = F->getFunctionType();
+    std::map<unsigned, Type *> ChangedType;
+    oclGetMutatedArgumentTypesByBuiltin(FT, ChangedType, F);
+    mutateFuncArgType(ChangedType, F);
+  }
+
+  // SPIR-V logical layout requires all function declarations go before
+  // function definitions.
+  std::vector<Function *> Decls, Defs;
+  for (Module::iterator I = M->begin(), E = M->end(); I != E; ++I) {
+    if (isBuiltinTransToInst(&*I) || isBuiltinTransToExtInst(&*I)
+        || I->getName().startswith(SPCV_CAST) ||
+        I->getName().startswith(LLVM_MEMCPY))
+      continue;
+    if (I->isDeclaration())
+      Decls.push_back(&*I);
+    else
+      Defs.push_back(&*I);
+  }
+  for (auto I:Decls)
+    transFunctionDecl(&*I);
+  for (auto I:Defs)
+    transFunction(&*I);
+
+  if (!transOCLKernelMetadata())
+    return false;
+  if (!transExecutionMode())
+    return false;
+
+  BM->optimizeDecorates();
+  BM->resolveUnknownStructFields();
+  BM->createForwardPointers();
+  return true;
+}
+
+llvm::IntegerType* LLVMToSPIRV::getSizetType() {
+  return IntegerType::getIntNTy(M->getContext(),
+    M->getDataLayout().getPointerSizeInBits());
+}
+
+void
+LLVMToSPIRV::oclGetMutatedArgumentTypesByBuiltin(
+    llvm::FunctionType* FT, std::map<unsigned, Type*>& ChangedType,
+    Function* F) {
+  auto Name = F->getName();
+  std::string Demangled;
+  if (!oclIsBuiltin(Name, &Demangled))
+    return;
+  if (Demangled.find(kSPIRVName::SampledImage) == std::string::npos)
+    return;
+  if (FT->getParamType(1)->isIntegerTy())
+    ChangedType[1] = getSamplerType(F->getParent());
+}
+
+SPIRVInstruction *
+LLVMToSPIRV::transBuiltinToInst(const std::string& DemangledName,
+    const std::string &MangledName, CallInst* CI, SPIRVBasicBlock* BB) {
+  SmallVector<std::string, 2> Dec;
+  auto OC = getSPIRVFuncOC(DemangledName, &Dec);
+
+  if (OC == OpNop)
+    return nullptr;
+
+  auto Inst = transBuiltinToInstWithoutDecoration(OC, CI, BB);
+  addDecorations(Inst, Dec);
+  return Inst;
+}
+
+bool
+LLVMToSPIRV::transExecutionMode() {
+  if (auto NMD = SPIRVMDWalker(*M).getNamedMD(kSPIRVMD::ExecutionMode)) {
+    while (!NMD.atEnd()) {
+      unsigned EMode = ~0U;
+      unsigned EModel = ~0U;
+      Function *F = nullptr;
+      auto N = NMD.nextOp(); /* execution mode MDNode */
+      N.nextOp() /* entry point MDNode */
+          .get(EModel)
+          .get(F)
+          .done()
+       .get(EMode);
+      assert (EModel == spv::ExecutionModelKernel &&
+          "Unsupported execution model");
+      SPIRVFunction *BF = static_cast<SPIRVFunction *>(getTranslatedValue(F));
+      assert(BF && "Invalid kernel function");
+      switch(EMode) {
+      case spv::ExecutionModeContractionOff:
+        BF->addExecutionMode(new SPIRVExecutionMode(BF,
+            ExecutionModeContractionOff));
+        break;
+      case spv::ExecutionModeLocalSize: {
+        unsigned X, Y, Z;
+        N.get(X).get(Y).get(Z);
+        BF->addExecutionMode(new SPIRVExecutionMode(BF,
+            ExecutionModeLocalSize, X, Y, Z));
+      }
+      break;
+      case spv::ExecutionModeLocalSizeHint: {
+        unsigned X, Y, Z;
+        N.get(X).get(Y).get(Z);
+        BF->addExecutionMode(new SPIRVExecutionMode(BF,
+            ExecutionModeLocalSizeHint, X, Y, Z));
+      }
+      break;
+      case spv::ExecutionModeVecTypeHint: {
+        unsigned X;
+        N.get(X);
+        BF->addExecutionMode(new SPIRVExecutionMode(BF,
+            ExecutionModeVecTypeHint, X));
+      }
+      break;
+      default:
+        llvm_unreachable("invalid execution mode");
+      }
+    }
+  }
+  return true;
+}
+
+bool
+LLVMToSPIRV::transOCLKernelMetadata() {
+  NamedMDNode *KernelMDs = M->getNamedMetadata(SPIR_MD_KERNELS);
+  std::vector<std::string> argAccessQual;
+  if (!KernelMDs)
+    return true;
+
+  for (unsigned I = 0, E = KernelMDs->getNumOperands(); I < E; ++I) {
+    MDNode *KernelMD = KernelMDs->getOperand(I);
+    if (KernelMD->getNumOperands() == 0)
+      continue;
+    Function *Kernel = mdconst::dyn_extract<Function>(KernelMD->getOperand(0));
+
+    SPIRVFunction *BF = static_cast<SPIRVFunction *>(getTranslatedValue(Kernel));
+    assert(BF && "Kernel function should be translated first");
+    assert(Kernel && oclIsKernel(Kernel)
+            && "Invalid kernel calling convention or metadata");
+    for (unsigned MI = 1, ME = KernelMD->getNumOperands(); MI < ME; ++MI) {
+      MDNode *MD = dyn_cast<MDNode>(KernelMD->getOperand(MI));
+      if (!MD)
+        continue;
+      MDString *NameMD = dyn_cast<MDString>(MD->getOperand(0));
+      if (!NameMD)
+        continue;
+      StringRef Name = NameMD->getString();
+      if (Name == SPIR_MD_KERNEL_ARG_TYPE_QUAL) {
+        foreachKernelArgMD(MD, BF,
+            [](const std::string &Str, SPIRVFunctionParameter *BA){
+          if (Str.find("volatile") != std::string::npos)
+            BA->addDecorate(new SPIRVDecorate(DecorationVolatile, BA));
+          if (Str.find("restrict") != std::string::npos)
+            BA->addDecorate(new SPIRVDecorate(DecorationFuncParamAttr,
+                BA, FunctionParameterAttributeNoAlias));
+          if (Str.find("const") != std::string::npos)
+            BA->addDecorate(new SPIRVDecorate(DecorationFuncParamAttr,
+                BA, FunctionParameterAttributeNoWrite));
+          });
+      } else if (Name == SPIR_MD_KERNEL_ARG_NAME) {
+        foreachKernelArgMD(MD, BF,
+            [=](const std::string &Str, SPIRVFunctionParameter *BA){
+            BM->setName(BA, Str);
+          });
+      }
+    }
+  }
+  return true;
+}
+
+bool
+LLVMToSPIRV::transSourceLanguage() {
+  auto Src = getSPIRVSource(M);
+  SrcLang = std::get<0>(Src);
+  SrcLangVer = std::get<1>(Src);
+  BM->setSourceLanguage(static_cast<SourceLanguage>(SrcLang), SrcLangVer);
+  return true;
+}
+
+bool
+LLVMToSPIRV::transExtension() {
+  if (auto N = SPIRVMDWalker(*M).getNamedMD(kSPIRVMD::Extension)) {
+    while (!N.atEnd()) {
+      std::string S;
+      N.nextOp().get(S);
+      assert(!S.empty() && "Invalid extension");
+      BM->getExtension().insert(S);
+    }
+  }
+  if (auto N = SPIRVMDWalker(*M).getNamedMD(kSPIRVMD::SourceExtension)) {
+    while (!N.atEnd()) {
+      std::string S;
+      N.nextOp().get(S);
+      assert(!S.empty() && "Invalid extension");
+      BM->getSourceExtension().insert(S);
+    }
+  }
+  for (auto &I:map<SPIRVCapabilityKind>(rmap<OclExt::Kind>(BM->getExtension())))
+    BM->addCapability(I);
+
+  return true;
+}
+
+void
+LLVMToSPIRV::dumpUsers(Value* V) {
+  SPIRVDBG(dbgs() << "Users of " << *V << " :\n");
+  for (auto UI = V->user_begin(), UE = V->user_end();
+      UI != UE; ++UI)
+    SPIRVDBG(dbgs() << "  " << **UI << '\n');
+}
+
+Op
+LLVMToSPIRV::transBoolOpCode(SPIRVValue* Opn, Op OC) {
+  if (!Opn->getType()->isTypeVectorOrScalarBool())
+    return OC;
+  IntBoolOpMap::find(OC, &OC);
+  return OC;
+}
+
+SPIRVInstruction *
+LLVMToSPIRV::transBuiltinToInstWithoutDecoration(Op OC,
+    CallInst* CI, SPIRVBasicBlock* BB) {
+  if (isGroupOpCode(OC))
+    BM->addCapability(CapabilityGroups);
+  switch (OC) {
+  case OpControlBarrier: {
+    auto BArgs = transValue(getArguments(CI), BB);
+    return BM->addControlBarrierInst(
+      BArgs[0], BArgs[1], BArgs[2], BB);
+    }
+    break;
+  case OpGroupAsyncCopy: {
+    auto BArgs = transValue(getArguments(CI), BB);
+    return BM->addAsyncGroupCopy(BArgs[0], BArgs[1], BArgs[2], BArgs[3],
+                                 BArgs[4], BArgs[5], BB);
+    }
+    break;
+  default: {
+    if (isCvtOpCode(OC) && OC != OpGenericCastToPtrExplicit) {
+      return BM->addUnaryInst(OC, transType(CI->getType()),
+        transValue(CI->getArgOperand(0), BB), BB);
+    } else if (isCmpOpCode(OC)) {
+      assert(CI && CI->getNumArgOperands() == 2 && "Invalid call inst");
+      auto ResultTy = CI->getType();
+      Type *BoolTy = IntegerType::getInt1Ty(M->getContext());
+      auto IsVector = ResultTy->isVectorTy();
+      if (IsVector)
+        BoolTy = VectorType::get(BoolTy, ResultTy->getVectorNumElements());
+      auto BBT = transType(BoolTy);
+      auto Cmp = BM->addCmpInst(OC, BBT,
+        transValue(CI->getArgOperand(0), BB),
+        transValue(CI->getArgOperand(1), BB), BB);
+      auto Zero = transValue(Constant::getNullValue(ResultTy), BB);
+      auto One = transValue(
+          IsVector ? Constant::getAllOnesValue(ResultTy) : getInt32(M, 1), BB);
+      return BM->addSelectInst(Cmp, One, Zero, BB);
+    } else if (isBinaryOpCode(OC)) {
+      assert(CI && CI->getNumArgOperands() == 2 && "Invalid call inst");
+      return BM->addBinaryInst(OC, transType(CI->getType()),
+        transValue(CI->getArgOperand(0), BB),
+        transValue(CI->getArgOperand(1), BB), BB);
+    } else if (CI->getNumArgOperands() == 1 &&
+        !CI->getType()->isVoidTy() &&
+        !hasExecScope(OC) &&
+        !isAtomicOpCode(OC)) {
+      return BM->addUnaryInst(OC, transType(CI->getType()),
+        transValue(CI->getArgOperand(0), BB), BB);
+    } else {
+      auto Args = getArguments(CI);
+      SPIRVType *SPRetTy = nullptr;
+      Type *RetTy = CI->getType();
+      auto F = CI->getCalledFunction();
+      if (!RetTy->isVoidTy()) {
+        SPRetTy = transType(RetTy);
+      } else if (Args.size() > 0 && F->arg_begin()->hasStructRetAttr()) {
+        SPRetTy = transType(F->arg_begin()->getType()->getPointerElementType());
+        Args.erase(Args.begin());
+      }
+      auto SPI = BM->addInstTemplate(OC, BB, SPRetTy);
+      std::vector<SPIRVWord> SPArgs;
+      for (size_t I = 0, E = Args.size(); I != E; ++I) {
+        assert((!isFunctionPointerType(Args[I]->getType()) ||
+               isa<Function>(Args[I])) &&
+               "Invalid function pointer argument");
+        SPArgs.push_back(SPI->isOperandLiteral(I) ?
+            cast<ConstantInt>(Args[I])->getZExtValue() :
+            transValue(Args[I], BB)->getId());
+      }
+      SPI->setOpWordsAndValidate(SPArgs);
+      if (!SPRetTy || !SPRetTy->isTypeStruct())
+        return SPI;
+      std::vector<SPIRVWord> Mem;
+      SPIRVDBG(spvdbgs() << *SPI << '\n');
+      return BM->addStoreInst(transValue(CI->getArgOperand(0), BB), SPI,
+          Mem, BB);
+    }
+  }
+  }
+  return nullptr;
+}
+
+
+SPIRVId
+LLVMToSPIRV::addInt32(int I) {
+  return transValue(getInt32(M, I), nullptr, false)->getId();
+}
+
+SPIRV::SPIRVLinkageTypeKind
+LLVMToSPIRV::transLinkageType(const GlobalValue* GV) {
+  if(GV->isDeclarationForLinker())
+    return SPIRVLinkageTypeKind::LinkageTypeImport;
+  if(GV->hasInternalLinkage() || GV->hasPrivateLinkage())
+    return SPIRVLinkageTypeKind::LinkageTypeInternal;
+  return SPIRVLinkageTypeKind::LinkageTypeExport;
+}
+} // end of SPIRV namespace
+
+char LLVMToSPIRV::ID = 0;
+
+INITIALIZE_PASS_BEGIN(LLVMToSPIRV, "llvmtospv", "Translate LLVM to SPIR-V",
+    false, false)
+INITIALIZE_PASS_DEPENDENCY(OCLTypeToSPIRV)
+INITIALIZE_PASS_END(LLVMToSPIRV, "llvmtospv", "Translate LLVM to SPIR-V",
+    false, false)
+
+ModulePass *llvm::createLLVMToSPIRV(SPIRVModule *SMod) {
+  return new LLVMToSPIRV(SMod);
+}
+
+void
+addPassesForSPIRV(legacy::PassManager &PassMgr) {
+  if (SPIRVMemToReg)
+    PassMgr.add(createPromoteMemoryToRegisterPass());
+  PassMgr.add(createTransOCLMD());
+  PassMgr.add(createOCL21ToSPIRV());
+  PassMgr.add(createSPIRVLowerOCLBlocks());
+  PassMgr.add(createOCLTypeToSPIRV());
+  PassMgr.add(createOCL20ToSPIRV());
+  PassMgr.add(createSPIRVRegularizeLLVM());
+  PassMgr.add(createSPIRVLowerConstExpr());
+  PassMgr.add(createSPIRVLowerBool());
+}
+
+bool
+llvm::WriteSPIRV(Module *M, llvm::raw_ostream &OS, std::string &ErrMsg) {
+  std::unique_ptr<SPIRVModule> BM(SPIRVModule::createSPIRVModule());
+  legacy::PassManager PassMgr;
+  addPassesForSPIRV(PassMgr);
+  PassMgr.add(createLLVMToSPIRV(BM.get()));
+  PassMgr.run(*M);
+
+  if (BM->getError(ErrMsg) != SPIRVEC_Success)
+    return false;
+  OS << *BM;
+  return true;
+}
+
+bool
+llvm::RegularizeLLVMForSPIRV(Module *M, std::string &ErrMsg) {
+  std::unique_ptr<SPIRVModule> BM(SPIRVModule::createSPIRVModule());
+  legacy::PassManager PassMgr;
+  addPassesForSPIRV(PassMgr);
+  PassMgr.run(*M);
+  return true;
+}
+
diff --git a/SPIRVWriterPass.cpp b/SPIRVWriterPass.cpp
index 071c26c..6042ff3 100644
--- a/SPIRVWriterPass.cpp
+++ b/SPIRVWriterPass.cpp
@@ -33,7 +33,7 @@ namespace {
     explicit WriteSPIRVPass(raw_ostream &o)
       : ModulePass(ID), OS(o) {}
 
-    const char *getPassName() const override { return "SPIRV Writer"; }
+    StringRef getPassName() const override { return "SPIRV Writer"; }
 
     bool runOnModule(Module &M) override {
       // FIXME: at the moment LLVM/SPIR-V translation errors are ignored.
diff --git a/libSPIRV/SPIRVType.h b/libSPIRV/SPIRVType.h
index 0f4ee02..c093805 100644
--- a/libSPIRV/SPIRVType.h
+++ b/libSPIRV/SPIRVType.h
@@ -301,7 +301,7 @@ class SPIRVTypeArray:public SPIRVType {
   SPIRVType *getElementType() const { return ElemType;}
   SPIRVConstant *getLength() const;
   SPIRVCapVec getRequiredCapability() const {
-    return std::move(getElementType()->getRequiredCapability());
+    return getElementType()->getRequiredCapability();
   }
   virtual std::vector<SPIRVEntry*> getNonLiteralOperands() const {
     std::vector<SPIRVEntry*> Operands(2, ElemType);
diff --git a/libSPIRV/SPIRVValue.h b/libSPIRV/SPIRVValue.h
index 5c9dcee..c87e41c 100644
--- a/libSPIRV/SPIRVValue.h
+++ b/libSPIRV/SPIRVValue.h
@@ -1,364 +1,364 @@
-//===- SPIRVValue.h - Class to represent a SPIR-V Value ----------*- C++ -*-===//
-//
-//                     The LLVM/SPIRV Translator
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-// Copyright (c) 2014 Advanced Micro Devices, Inc. All rights reserved.
-//
-// Permission is hereby granted, free of charge, to any person obtaining a
-// copy of this software and associated documentation files (the "Software"),
-// to deal with the Software without restriction, including without limitation
-// the rights to use, copy, modify, merge, publish, distribute, sublicense,
-// and/or sell copies of the Software, and to permit persons to whom the
-// Software is furnished to do so, subject to the following conditions:
-//
-// Redistributions of source code must retain the above copyright notice,
-// this list of conditions and the following disclaimers.
-// Redistributions in binary form must reproduce the above copyright notice,
-// this list of conditions and the following disclaimers in the documentation
-// and/or other materials provided with the distribution.
-// Neither the names of Advanced Micro Devices, Inc., nor the names of its
-// contributors may be used to endorse or promote products derived from this
-// Software without specific prior written permission.
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-// CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
-// THE SOFTWARE.
-//
-//===----------------------------------------------------------------------===//
-/// \file
-///
-/// This file defines the values defined in SPIR-V spec with op codes.
-///
-/// The name of the SPIR-V values follow the op code name in the spec.
-/// This is for readability and ease of using macro to handle types.
-//
-//===----------------------------------------------------------------------===//
-
-#ifndef SPIRVVALUE_HPP_
-#define SPIRVVALUE_HPP_
-
-#include "SPIRVEntry.h"
-#include "SPIRVType.h"
-#include "SPIRVDecorate.h"
-
-#include <iostream>
-#include <map>
-#include <memory>
-
-namespace SPIRV{
-
-class SPIRVValue: public SPIRVEntry {
-public:
-  // Complete constructor for value with id and type
-  SPIRVValue(SPIRVModule *M, unsigned TheWordCount, Op TheOpCode,
-      SPIRVType *TheType, SPIRVId TheId)
-    :SPIRVEntry(M, TheWordCount, TheOpCode, TheId), Type(TheType) {
-    validate();
-  }
-  // Complete constructor for value with type but without id
-  SPIRVValue(SPIRVModule *M, unsigned TheWordCount, Op TheOpCode,
-      SPIRVType *TheType)
-    :SPIRVEntry(M, TheWordCount, TheOpCode), Type(TheType) {
-    setHasNoId();
-    validate();
-  }
-  // Complete constructor for value with id but without type
-  SPIRVValue(SPIRVModule *M, unsigned TheWordCount, Op TheOpCode,
-      SPIRVId TheId)
-    :SPIRVEntry(M, TheWordCount, TheOpCode, TheId), Type(NULL) {
-    setHasNoType();
-    validate();
-  }
-  // Complete constructor for value without id and type
-  SPIRVValue(SPIRVModule *M, unsigned TheWordCount, Op TheOpCode)
-    :SPIRVEntry(M, TheWordCount, TheOpCode), Type(NULL) {
-    setHasNoId();
-    setHasNoType();
-    validate();
-  }
-  // Incomplete constructor
-  SPIRVValue(Op TheOpCode):SPIRVEntry(TheOpCode), Type(NULL) {}
-
-  bool hasType()const { return !(Attrib & SPIRVEA_NOTYPE);}
-  SPIRVType *getType()const {
-    assert(hasType() && "value has no type");
-    return Type;
-  }
-  bool isVolatile()const;
-  bool hasAlignment(SPIRVWord *Result=0)const;
-
-  void setAlignment(SPIRVWord);
-  void setVolatile(bool IsVolatile);
-
-  void validate()const {
-    SPIRVEntry::validate();
-    assert((!hasType() || Type) && "Invalid type");
-  }
-
-  void setType(SPIRVType *Ty) {
-    Type = Ty;
-    assert(!Ty || !Ty->isTypeVoid() || OpCode == OpFunction);
-    if (Ty && (!Ty->isTypeVoid() || OpCode == OpFunction))
-      setHasType();
-    else
-      setHasNoType();
-  }
-
-  SPIRVCapVec getRequiredCapability() const {
-    SPIRVCapVec CV;
-    if (!hasType())
-      return std::move(CV);
-    return std::move(Type->getRequiredCapability());
-  }
-
-protected:
-  void setHasNoType() { Attrib |= SPIRVEA_NOTYPE;}
-  void setHasType() { Attrib &= ~SPIRVEA_NOTYPE;}
-
-  SPIRVType *Type;                 // Value Type
-};
-
-class SPIRVConstant: public SPIRVValue {
-public:
-  // Complete constructor for integer constant
-  SPIRVConstant(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId,
-      uint64_t TheValue)
-    :SPIRVValue(M, 0, OpConstant, TheType, TheId){
-    Union.UInt64Val = TheValue;
-    recalculateWordCount();
-    validate();
-  }
-  // Complete constructor for float constant
-  SPIRVConstant(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId, float TheValue)
-    :SPIRVValue(M, 0, OpConstant, TheType, TheId){
-    Union.FloatVal = TheValue;
-    recalculateWordCount();
-    validate();
-  }
-  // Complete constructor for double constant
-  SPIRVConstant(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId, double TheValue)
-    :SPIRVValue(M, 0, OpConstant, TheType, TheId){
-    Union.DoubleVal = TheValue;
-    recalculateWordCount();
-    validate();
-  }
-  // Incomplete constructor
-  SPIRVConstant():SPIRVValue(OpConstant), NumWords(0){}
-  uint64_t getZExtIntValue() const { return Union.UInt64Val;}
-  float getFloatValue() const { return Union.FloatVal;}
-  double getDoubleValue() const { return Union.DoubleVal;}
-protected:
-  void recalculateWordCount() {
-    NumWords = Type->getBitWidth()/32;
-    if (NumWords < 1)
-      NumWords = 1;
-    WordCount = 3 + NumWords;
-  }
-  void validate() const {
-    SPIRVValue::validate();
-    assert(NumWords >= 1 && NumWords <= 2 && "Invalid constant size");
-  }
-  void encode(spv_ostream &O) const {
-    getEncoder(O) << Type << Id;
-    for (unsigned i = 0; i < NumWords; ++i)
-      getEncoder(O) << Union.Words[i];
-  }
-  void setWordCount(SPIRVWord WordCount) {
-    SPIRVValue::setWordCount(WordCount);
-    NumWords = WordCount - 3;
-  }
-  void decode(std::istream &I) {
-    getDecoder(I) >> Type >> Id;
-    for (unsigned i = 0; i < NumWords; ++i)
-      getDecoder(I) >> Union.Words[i];
-  }
-
-  unsigned NumWords;
-  union UnionType{
-    uint64_t UInt64Val;
-    float FloatVal;
-    double DoubleVal;
-    SPIRVWord Words[2];
-    UnionType() {
-      UInt64Val = 0;
-    }
-  } Union;
-};
-
-template<Op OC>
-class SPIRVConstantEmpty: public SPIRVValue {
-public:
-  // Complete constructor
-  SPIRVConstantEmpty(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId)
-    :SPIRVValue(M, 3, OC, TheType, TheId){
-    validate();
-  }
-  // Incomplete constructor
-  SPIRVConstantEmpty():SPIRVValue(OC){}
-protected:
-  void validate() const {
-    SPIRVValue::validate();
-  }
-  _SPIRV_DEF_ENCDEC2(Type, Id)
-};
-
-template<Op OC>
-class SPIRVConstantBool: public SPIRVConstantEmpty<OC> {
-public:
-  // Complete constructor
-  SPIRVConstantBool(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId)
-    :SPIRVConstantEmpty<OC>(M, TheType, TheId){}
-  // Incomplete constructor
-  SPIRVConstantBool(){}
-protected:
-  void validate() const {
-    SPIRVConstantEmpty<OC>::validate();
-    assert(this->Type->isTypeBool() && "Invalid type");
-  }
-};
-
-typedef SPIRVConstantBool<OpConstantTrue> SPIRVConstantTrue;
-typedef SPIRVConstantBool<OpConstantFalse> SPIRVConstantFalse;
-
-class SPIRVConstantNull:
-    public SPIRVConstantEmpty<OpConstantNull> {
-public:
-  // Complete constructor
-  SPIRVConstantNull(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId)
-    :SPIRVConstantEmpty(M, TheType, TheId){
-    validate();
-  }
-  // Incomplete constructor
-  SPIRVConstantNull(){}
-protected:
-  void validate() const {
-    SPIRVConstantEmpty::validate();
-    assert((Type->isTypeComposite() ||
-            Type->isTypeOpaque() ||
-            Type->isTypeEvent() ||
-            Type->isTypePointer() ||
-            Type->isTypeReserveId() ||
-            Type->isTypeDeviceEvent()) &&
-            "Invalid type");
-  }
-};
-
-class SPIRVUndef:
-    public SPIRVConstantEmpty<OpUndef> {
-public:
-  // Complete constructor
-  SPIRVUndef(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId)
-    :SPIRVConstantEmpty(M, TheType, TheId){
-    validate();
-  }
-  // Incomplete constructor
-  SPIRVUndef(){}
-protected:
-  void validate() const {
-    SPIRVConstantEmpty::validate();
-  }
-};
-
-class SPIRVConstantComposite: public SPIRVValue {
-public:
-  // Complete constructor for composite constant
-  SPIRVConstantComposite(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId,
-      const std::vector<SPIRVValue *> TheElements)
-    :SPIRVValue(M, TheElements.size()+3, OpConstantComposite, TheType,
-        TheId){
-    Elements = getIds(TheElements);
-    validate();
-  }
-  // Incomplete constructor
-  SPIRVConstantComposite():SPIRVValue(OpConstantComposite){}
-  std::vector<SPIRVValue*> getElements()const {
-    return getValues(Elements);
-  }
-  std::vector<SPIRVEntry*> getNonLiteralOperands() const {
-    std::vector<SPIRVValue*> Elements = getElements();
-    return std::vector<SPIRVEntry*>(Elements.begin(), Elements.end());
-  }
-protected:
-  void validate() const {
-    SPIRVValue::validate();
-    for (auto &I:Elements)
-      getValue(I)->validate();
-  }
-  void setWordCount(SPIRVWord WordCount) {
-    Elements.resize(WordCount - 3);
-  }
-  _SPIRV_DEF_ENCDEC3(Type, Id, Elements)
-
-  std::vector<SPIRVId> Elements;
-};
-
-class SPIRVConstantSampler: public SPIRVValue {
-public:
-  const static Op OC = OpConstantSampler;
-  const static SPIRVWord WC = 6;
-  // Complete constructor
-  SPIRVConstantSampler(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId,
-      SPIRVWord TheAddrMode, SPIRVWord TheNormalized, SPIRVWord TheFilterMode)
-    :SPIRVValue(M, WC, OC, TheType, TheId), AddrMode(TheAddrMode),
-     Normalized(TheNormalized), FilterMode(TheFilterMode){
-    validate();
-  }
-  // Incomplete constructor
-  SPIRVConstantSampler():SPIRVValue(OC), AddrMode(SPIRVSAM_Invalid),
-      Normalized(SPIRVWORD_MAX), FilterMode(SPIRVSFM_Invalid){}
-
-  SPIRVWord getAddrMode() const {
-    return AddrMode;
-  }
-
-  SPIRVWord getFilterMode() const {
-    return FilterMode;
-  }
-
-  SPIRVWord getNormalized() const {
-    return Normalized;
-  }
-  SPIRVCapVec getRequiredCapability() const {
-    return getVec(CapabilityLiteralSampler);
-  }
-protected:
-  SPIRVWord AddrMode;
-  SPIRVWord Normalized;
-  SPIRVWord FilterMode;
-  void validate() const {
-    SPIRVValue::validate();
-    assert(OpCode == OC);
-    assert(WordCount == WC);
-    assert(Type->isTypeSampler());
-  }
-  _SPIRV_DEF_ENCDEC5(Type, Id, AddrMode, Normalized, FilterMode)
-};
-
-class SPIRVForward:public SPIRVValue, public SPIRVComponentExecutionModes {
-public:
-  const static Op OC = OpForward;
-  // Complete constructor
-  SPIRVForward(SPIRVModule *TheModule, SPIRVType *TheTy, SPIRVId TheId):
-    SPIRVValue(TheModule, 0, OC, TheId){
-    if (TheTy)
-      setType(TheTy);
-  }
-  SPIRVForward():SPIRVValue(OC) {
-    assert(0 && "should never be called");
-  }
-  _SPIRV_DEF_ENCDEC1(Id)
-  friend class SPIRVFunction;
-protected:
-  void validate() const {}
-};
-
-}
-
-
-#endif /* SPIRVVALUE_HPP_ */
+//===- SPIRVValue.h - Class to represent a SPIR-V Value ----------*- C++ -*-===//
+//
+//                     The LLVM/SPIRV Translator
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+// Copyright (c) 2014 Advanced Micro Devices, Inc. All rights reserved.
+//
+// Permission is hereby granted, free of charge, to any person obtaining a
+// copy of this software and associated documentation files (the "Software"),
+// to deal with the Software without restriction, including without limitation
+// the rights to use, copy, modify, merge, publish, distribute, sublicense,
+// and/or sell copies of the Software, and to permit persons to whom the
+// Software is furnished to do so, subject to the following conditions:
+//
+// Redistributions of source code must retain the above copyright notice,
+// this list of conditions and the following disclaimers.
+// Redistributions in binary form must reproduce the above copyright notice,
+// this list of conditions and the following disclaimers in the documentation
+// and/or other materials provided with the distribution.
+// Neither the names of Advanced Micro Devices, Inc., nor the names of its
+// contributors may be used to endorse or promote products derived from this
+// Software without specific prior written permission.
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+// CONTRIBUTORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS WITH
+// THE SOFTWARE.
+//
+//===----------------------------------------------------------------------===//
+/// \file
+///
+/// This file defines the values defined in SPIR-V spec with op codes.
+///
+/// The name of the SPIR-V values follow the op code name in the spec.
+/// This is for readability and ease of using macro to handle types.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef SPIRVVALUE_HPP_
+#define SPIRVVALUE_HPP_
+
+#include "SPIRVEntry.h"
+#include "SPIRVType.h"
+#include "SPIRVDecorate.h"
+
+#include <iostream>
+#include <map>
+#include <memory>
+
+namespace SPIRV{
+
+class SPIRVValue: public SPIRVEntry {
+public:
+  // Complete constructor for value with id and type
+  SPIRVValue(SPIRVModule *M, unsigned TheWordCount, Op TheOpCode,
+      SPIRVType *TheType, SPIRVId TheId)
+    :SPIRVEntry(M, TheWordCount, TheOpCode, TheId), Type(TheType) {
+    validate();
+  }
+  // Complete constructor for value with type but without id
+  SPIRVValue(SPIRVModule *M, unsigned TheWordCount, Op TheOpCode,
+      SPIRVType *TheType)
+    :SPIRVEntry(M, TheWordCount, TheOpCode), Type(TheType) {
+    setHasNoId();
+    validate();
+  }
+  // Complete constructor for value with id but without type
+  SPIRVValue(SPIRVModule *M, unsigned TheWordCount, Op TheOpCode,
+      SPIRVId TheId)
+    :SPIRVEntry(M, TheWordCount, TheOpCode, TheId), Type(NULL) {
+    setHasNoType();
+    validate();
+  }
+  // Complete constructor for value without id and type
+  SPIRVValue(SPIRVModule *M, unsigned TheWordCount, Op TheOpCode)
+    :SPIRVEntry(M, TheWordCount, TheOpCode), Type(NULL) {
+    setHasNoId();
+    setHasNoType();
+    validate();
+  }
+  // Incomplete constructor
+  SPIRVValue(Op TheOpCode):SPIRVEntry(TheOpCode), Type(NULL) {}
+
+  bool hasType()const { return !(Attrib & SPIRVEA_NOTYPE);}
+  SPIRVType *getType()const {
+    assert(hasType() && "value has no type");
+    return Type;
+  }
+  bool isVolatile()const;
+  bool hasAlignment(SPIRVWord *Result=0)const;
+
+  void setAlignment(SPIRVWord);
+  void setVolatile(bool IsVolatile);
+
+  void validate()const {
+    SPIRVEntry::validate();
+    assert((!hasType() || Type) && "Invalid type");
+  }
+
+  void setType(SPIRVType *Ty) {
+    Type = Ty;
+    assert(!Ty || !Ty->isTypeVoid() || OpCode == OpFunction);
+    if (Ty && (!Ty->isTypeVoid() || OpCode == OpFunction))
+      setHasType();
+    else
+      setHasNoType();
+  }
+
+  SPIRVCapVec getRequiredCapability() const {
+    SPIRVCapVec CV;
+    if (!hasType())
+      return CV;
+    return Type->getRequiredCapability();
+  }
+
+protected:
+  void setHasNoType() { Attrib |= SPIRVEA_NOTYPE;}
+  void setHasType() { Attrib &= ~SPIRVEA_NOTYPE;}
+
+  SPIRVType *Type;                 // Value Type
+};
+
+class SPIRVConstant: public SPIRVValue {
+public:
+  // Complete constructor for integer constant
+  SPIRVConstant(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId,
+      uint64_t TheValue)
+    :SPIRVValue(M, 0, OpConstant, TheType, TheId){
+    Union.UInt64Val = TheValue;
+    recalculateWordCount();
+    validate();
+  }
+  // Complete constructor for float constant
+  SPIRVConstant(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId, float TheValue)
+    :SPIRVValue(M, 0, OpConstant, TheType, TheId){
+    Union.FloatVal = TheValue;
+    recalculateWordCount();
+    validate();
+  }
+  // Complete constructor for double constant
+  SPIRVConstant(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId, double TheValue)
+    :SPIRVValue(M, 0, OpConstant, TheType, TheId){
+    Union.DoubleVal = TheValue;
+    recalculateWordCount();
+    validate();
+  }
+  // Incomplete constructor
+  SPIRVConstant():SPIRVValue(OpConstant), NumWords(0){}
+  uint64_t getZExtIntValue() const { return Union.UInt64Val;}
+  float getFloatValue() const { return Union.FloatVal;}
+  double getDoubleValue() const { return Union.DoubleVal;}
+protected:
+  void recalculateWordCount() {
+    NumWords = Type->getBitWidth()/32;
+    if (NumWords < 1)
+      NumWords = 1;
+    WordCount = 3 + NumWords;
+  }
+  void validate() const {
+    SPIRVValue::validate();
+    assert(NumWords >= 1 && NumWords <= 2 && "Invalid constant size");
+  }
+  void encode(spv_ostream &O) const {
+    getEncoder(O) << Type << Id;
+    for (unsigned i = 0; i < NumWords; ++i)
+      getEncoder(O) << Union.Words[i];
+  }
+  void setWordCount(SPIRVWord WordCount) {
+    SPIRVValue::setWordCount(WordCount);
+    NumWords = WordCount - 3;
+  }
+  void decode(std::istream &I) {
+    getDecoder(I) >> Type >> Id;
+    for (unsigned i = 0; i < NumWords; ++i)
+      getDecoder(I) >> Union.Words[i];
+  }
+
+  unsigned NumWords;
+  union UnionType{
+    uint64_t UInt64Val;
+    float FloatVal;
+    double DoubleVal;
+    SPIRVWord Words[2];
+    UnionType() {
+      UInt64Val = 0;
+    }
+  } Union;
+};
+
+template<Op OC>
+class SPIRVConstantEmpty: public SPIRVValue {
+public:
+  // Complete constructor
+  SPIRVConstantEmpty(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId)
+    :SPIRVValue(M, 3, OC, TheType, TheId){
+    validate();
+  }
+  // Incomplete constructor
+  SPIRVConstantEmpty():SPIRVValue(OC){}
+protected:
+  void validate() const {
+    SPIRVValue::validate();
+  }
+  _SPIRV_DEF_ENCDEC2(Type, Id)
+};
+
+template<Op OC>
+class SPIRVConstantBool: public SPIRVConstantEmpty<OC> {
+public:
+  // Complete constructor
+  SPIRVConstantBool(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId)
+    :SPIRVConstantEmpty<OC>(M, TheType, TheId){}
+  // Incomplete constructor
+  SPIRVConstantBool(){}
+protected:
+  void validate() const {
+    SPIRVConstantEmpty<OC>::validate();
+    assert(this->Type->isTypeBool() && "Invalid type");
+  }
+};
+
+typedef SPIRVConstantBool<OpConstantTrue> SPIRVConstantTrue;
+typedef SPIRVConstantBool<OpConstantFalse> SPIRVConstantFalse;
+
+class SPIRVConstantNull:
+    public SPIRVConstantEmpty<OpConstantNull> {
+public:
+  // Complete constructor
+  SPIRVConstantNull(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId)
+    :SPIRVConstantEmpty(M, TheType, TheId){
+    validate();
+  }
+  // Incomplete constructor
+  SPIRVConstantNull(){}
+protected:
+  void validate() const {
+    SPIRVConstantEmpty::validate();
+    assert((Type->isTypeComposite() ||
+            Type->isTypeOpaque() ||
+            Type->isTypeEvent() ||
+            Type->isTypePointer() ||
+            Type->isTypeReserveId() ||
+            Type->isTypeDeviceEvent()) &&
+            "Invalid type");
+  }
+};
+
+class SPIRVUndef:
+    public SPIRVConstantEmpty<OpUndef> {
+public:
+  // Complete constructor
+  SPIRVUndef(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId)
+    :SPIRVConstantEmpty(M, TheType, TheId){
+    validate();
+  }
+  // Incomplete constructor
+  SPIRVUndef(){}
+protected:
+  void validate() const {
+    SPIRVConstantEmpty::validate();
+  }
+};
+
+class SPIRVConstantComposite: public SPIRVValue {
+public:
+  // Complete constructor for composite constant
+  SPIRVConstantComposite(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId,
+      const std::vector<SPIRVValue *> TheElements)
+    :SPIRVValue(M, TheElements.size()+3, OpConstantComposite, TheType,
+        TheId){
+    Elements = getIds(TheElements);
+    validate();
+  }
+  // Incomplete constructor
+  SPIRVConstantComposite():SPIRVValue(OpConstantComposite){}
+  std::vector<SPIRVValue*> getElements()const {
+    return getValues(Elements);
+  }
+  std::vector<SPIRVEntry*> getNonLiteralOperands() const {
+    std::vector<SPIRVValue*> Elements = getElements();
+    return std::vector<SPIRVEntry*>(Elements.begin(), Elements.end());
+  }
+protected:
+  void validate() const {
+    SPIRVValue::validate();
+    for (auto &I:Elements)
+      getValue(I)->validate();
+  }
+  void setWordCount(SPIRVWord WordCount) {
+    Elements.resize(WordCount - 3);
+  }
+  _SPIRV_DEF_ENCDEC3(Type, Id, Elements)
+
+  std::vector<SPIRVId> Elements;
+};
+
+class SPIRVConstantSampler: public SPIRVValue {
+public:
+  const static Op OC = OpConstantSampler;
+  const static SPIRVWord WC = 6;
+  // Complete constructor
+  SPIRVConstantSampler(SPIRVModule *M, SPIRVType *TheType, SPIRVId TheId,
+      SPIRVWord TheAddrMode, SPIRVWord TheNormalized, SPIRVWord TheFilterMode)
+    :SPIRVValue(M, WC, OC, TheType, TheId), AddrMode(TheAddrMode),
+     Normalized(TheNormalized), FilterMode(TheFilterMode){
+    validate();
+  }
+  // Incomplete constructor
+  SPIRVConstantSampler():SPIRVValue(OC), AddrMode(SPIRVSAM_Invalid),
+      Normalized(SPIRVWORD_MAX), FilterMode(SPIRVSFM_Invalid){}
+
+  SPIRVWord getAddrMode() const {
+    return AddrMode;
+  }
+
+  SPIRVWord getFilterMode() const {
+    return FilterMode;
+  }
+
+  SPIRVWord getNormalized() const {
+    return Normalized;
+  }
+  SPIRVCapVec getRequiredCapability() const {
+    return getVec(CapabilityLiteralSampler);
+  }
+protected:
+  SPIRVWord AddrMode;
+  SPIRVWord Normalized;
+  SPIRVWord FilterMode;
+  void validate() const {
+    SPIRVValue::validate();
+    assert(OpCode == OC);
+    assert(WordCount == WC);
+    assert(Type->isTypeSampler());
+  }
+  _SPIRV_DEF_ENCDEC5(Type, Id, AddrMode, Normalized, FilterMode)
+};
+
+class SPIRVForward:public SPIRVValue, public SPIRVComponentExecutionModes {
+public:
+  const static Op OC = OpForward;
+  // Complete constructor
+  SPIRVForward(SPIRVModule *TheModule, SPIRVType *TheTy, SPIRVId TheId):
+    SPIRVValue(TheModule, 0, OC, TheId){
+    if (TheTy)
+      setType(TheTy);
+  }
+  SPIRVForward():SPIRVValue(OC) {
+    assert(0 && "should never be called");
+  }
+  _SPIRV_DEF_ENCDEC1(Id)
+  friend class SPIRVFunction;
+protected:
+  void validate() const {}
+};
+
+}
+
+
+#endif /* SPIRVVALUE_HPP_ */