Add R600 backend

A new backend supporting AMD GPUs: Radeon HD2XXX - HD7XXX

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@169915 91177308-0d34-0410-b5e6-96231b3b80d8

146 files changed, 20232 insertions, 1 deletions

diff --git a/include/llvm/Intrinsics.td b/include/llvm/Intrinsics.td
index d3a548cd4e..d5c01ccbf3 100644
--- a/include/llvm/Intrinsics.td
+++ b/include/llvm/Intrinsics.td
@@ -472,3 +472,4 @@ include "llvm/IntrinsicsXCore.td"
 include "llvm/IntrinsicsHexagon.td"
 include "llvm/IntrinsicsNVVM.td"
 include "llvm/IntrinsicsMips.td"
+include "llvm/IntrinsicsR600.td"
diff --git a/include/llvm/IntrinsicsR600.td b/include/llvm/IntrinsicsR600.td
new file mode 100644
index 0000000000..ecb5668d8e
--- /dev/null
+++ b/include/llvm/IntrinsicsR600.td
@@ -0,0 +1,36 @@
+//===- IntrinsicsR600.td - Defines R600 intrinsics ---------*- tablegen -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines all of the R600-specific intrinsics.
+//
+//===----------------------------------------------------------------------===//
+
+let TargetPrefix = "r600" in {
+
+class R600ReadPreloadRegisterIntrinsic<string name>
+  : Intrinsic<[llvm_i32_ty], [], [IntrNoMem]>,
+    GCCBuiltin<name>;
+
+multiclass R600ReadPreloadRegisterIntrinsic_xyz<string prefix> {
+  def _x : R600ReadPreloadRegisterIntrinsic<!strconcat(prefix, "_x")>;
+  def _y : R600ReadPreloadRegisterIntrinsic<!strconcat(prefix, "_y")>;
+  def _z : R600ReadPreloadRegisterIntrinsic<!strconcat(prefix, "_z")>;
+}
+
+defm int_r600_read_global_size : R600ReadPreloadRegisterIntrinsic_xyz <
+                                       "__builtin_r600_read_global_size">;
+defm int_r600_read_local_size : R600ReadPreloadRegisterIntrinsic_xyz <
+                                       "__builtin_r600_read_local_size">;
+defm int_r600_read_ngroups : R600ReadPreloadRegisterIntrinsic_xyz <
+                                       "__builtin_r600_read_ngroups">;
+defm int_r600_read_tgid : R600ReadPreloadRegisterIntrinsic_xyz <
+                                       "__builtin_r600_read_tgid">;
+defm int_r600_read_tidig : R600ReadPreloadRegisterIntrinsic_xyz <
+                                       "__builtin_r600_read_tidig">;
+} // End TargetPrefix = "r600"
diff --git a/lib/Target/LLVMBuild.txt b/lib/Target/LLVMBuild.txt
index eb6c779f45..f3a9c1c3e7 100644
--- a/lib/Target/LLVMBuild.txt
+++ b/lib/Target/LLVMBuild.txt
@@ -16,7 +16,7 @@
 ;===------------------------------------------------------------------------===;
 
 [common]
-subdirectories = ARM CppBackend Hexagon MBlaze MSP430 NVPTX Mips PowerPC Sparc X86 XCore
+subdirectories = ARM CppBackend Hexagon MBlaze MSP430 NVPTX Mips PowerPC R600 Sparc X86 XCore
 
 ; This is a special group whose required libraries are extended (by llvm-build)
 ; with the best execution engine (the native JIT, if available, or the
diff --git a/lib/Target/R600/AMDGPU.h b/lib/Target/R600/AMDGPU.h
new file mode 100644
index 0000000000..40864b09dd
--- /dev/null
+++ b/lib/Target/R600/AMDGPU.h
@@ -0,0 +1,48 @@
+//===-- AMDGPU.h - MachineFunction passes hw codegen --------------*- C++ -*-=//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//===----------------------------------------------------------------------===//
+
+#ifndef AMDGPU_H
+#define AMDGPU_H
+
+#include "AMDGPUTargetMachine.h"
+#include "llvm/Support/TargetRegistry.h"
+#include "llvm/Target/TargetMachine.h"
+
+namespace llvm {
+
+class FunctionPass;
+class AMDGPUTargetMachine;
+
+// R600 Passes
+FunctionPass* createR600KernelParametersPass(const DataLayout *TD);
+FunctionPass *createR600ExpandSpecialInstrsPass(TargetMachine &tm);
+
+// SI Passes
+FunctionPass *createSIAssignInterpRegsPass(TargetMachine &tm);
+FunctionPass *createSILowerControlFlowPass(TargetMachine &tm);
+FunctionPass *createSICodeEmitterPass(formatted_raw_ostream &OS);
+FunctionPass *createSILowerLiteralConstantsPass(TargetMachine &tm);
+FunctionPass *createSIFixSGPRLivenessPass(TargetMachine &tm);
+
+// Passes common to R600 and SI
+FunctionPass *createAMDGPUConvertToISAPass(TargetMachine &tm);
+
+} // End namespace llvm
+
+namespace ShaderType {
+  enum Type {
+    PIXEL = 0,
+    VERTEX = 1,
+    GEOMETRY = 2,
+    COMPUTE = 3
+  };
+}
+
+#endif // AMDGPU_H
diff --git a/lib/Target/R600/AMDGPU.td b/lib/Target/R600/AMDGPU.td
new file mode 100644
index 0000000000..40f474161a
--- /dev/null
+++ b/lib/Target/R600/AMDGPU.td
@@ -0,0 +1,40 @@
+//===-- AMDIL.td - AMDIL Tablegen files --*- tablegen -*-------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+
+// Include AMDIL TD files
+include "AMDILBase.td"
+
+
+def AMDGPUInstrInfo : InstrInfo {
+  let guessInstructionProperties = 1;
+}
+
+//===----------------------------------------------------------------------===//
+// Declare the target which we are implementing
+//===----------------------------------------------------------------------===//
+def AMDGPUAsmWriter : AsmWriter {
+    string AsmWriterClassName = "InstPrinter";
+    int Variant = 0;
+    bit isMCAsmWriter = 1;
+}
+
+def AMDGPU : Target {
+  // Pull in Instruction Info:
+  let InstructionSet = AMDGPUInstrInfo;
+  let AssemblyWriters = [AMDGPUAsmWriter];
+}
+
+// Include AMDGPU TD files
+include "R600Schedule.td"
+include "SISchedule.td"
+include "Processors.td"
+include "AMDGPUInstrInfo.td"
+include "AMDGPUIntrinsics.td"
+include "AMDGPURegisterInfo.td"
+include "AMDGPUInstructions.td"
diff --git a/lib/Target/R600/AMDGPUAsmPrinter.cpp b/lib/Target/R600/AMDGPUAsmPrinter.cpp
new file mode 100644
index 0000000000..4553c4556c
--- /dev/null
+++ b/lib/Target/R600/AMDGPUAsmPrinter.cpp
@@ -0,0 +1,138 @@
+//===-- AMDGPUAsmPrinter.cpp - AMDGPU Assebly printer  --------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+///
+/// The AMDGPUAsmPrinter is used to print both assembly string and also binary
+/// code.  When passed an MCAsmStreamer it prints assembly and when passed
+/// an MCObjectStreamer it outputs binary code.
+//
+//===----------------------------------------------------------------------===//
+//
+
+
+#include "AMDGPUAsmPrinter.h"
+#include "AMDGPU.h"
+#include "SIMachineFunctionInfo.h"
+#include "SIRegisterInfo.h"
+#include "llvm/MC/MCStreamer.h"
+#include "llvm/Target/TargetLoweringObjectFile.h"
+#include "llvm/Support/TargetRegistry.h"
+
+using namespace llvm;
+
+
+static AsmPrinter *createAMDGPUAsmPrinterPass(TargetMachine &tm,
+                                              MCStreamer &Streamer) {
+  return new AMDGPUAsmPrinter(tm, Streamer);
+}
+
+extern "C" void LLVMInitializeR600AsmPrinter() {
+  TargetRegistry::RegisterAsmPrinter(TheAMDGPUTarget, createAMDGPUAsmPrinterPass);
+}
+
+/// We need to override this function so we can avoid
+/// the call to EmitFunctionHeader(), which the MCPureStreamer can't handle.
+bool AMDGPUAsmPrinter::runOnMachineFunction(MachineFunction &MF) {
+  const AMDGPUSubtarget &STM = TM.getSubtarget<AMDGPUSubtarget>();
+  if (STM.dumpCode()) {
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
+    MF.dump();
+#endif
+  }
+  SetupMachineFunction(MF);
+  OutStreamer.SwitchSection(getObjFileLowering().getTextSection());
+  if (STM.device()->getGeneration() > AMDGPUDeviceInfo::HD6XXX) {
+    EmitProgramInfo(MF);
+  }
+  EmitFunctionBody();
+  return false;
+}
+
+void AMDGPUAsmPrinter::EmitProgramInfo(MachineFunction &MF) {
+  unsigned MaxSGPR = 0;
+  unsigned MaxVGPR = 0;
+  bool VCCUsed = false;
+  const SIRegisterInfo * RI =
+                static_cast<const SIRegisterInfo*>(TM.getRegisterInfo());
+
+  for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
+                                                  BB != BB_E; ++BB) {
+    MachineBasicBlock &MBB = *BB;
+    for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
+                                                    I != E; ++I) {
+      MachineInstr &MI = *I;
+
+      unsigned numOperands = MI.getNumOperands();
+      for (unsigned op_idx = 0; op_idx < numOperands; op_idx++) {
+        MachineOperand & MO = MI.getOperand(op_idx);
+        unsigned maxUsed;
+        unsigned width = 0;
+        bool isSGPR = false;
+        unsigned reg;
+        unsigned hwReg;
+        if (!MO.isReg()) {
+          continue;
+        }
+        reg = MO.getReg();
+        if (reg == AMDGPU::VCC) {
+          VCCUsed = true;
+          continue;
+        }
+        switch (reg) {
+        default: break;
+        case AMDGPU::EXEC:
+        case AMDGPU::SI_LITERAL_CONSTANT:
+        case AMDGPU::SREG_LIT_0:
+        case AMDGPU::M0:
+          continue;
+        }
+
+        if (AMDGPU::SReg_32RegClass.contains(reg)) {
+          isSGPR = true;
+          width = 1;
+        } else if (AMDGPU::VReg_32RegClass.contains(reg)) {
+          isSGPR = false;
+          width = 1;
+        } else if (AMDGPU::SReg_64RegClass.contains(reg)) {
+          isSGPR = true;
+          width = 2;
+        } else if (AMDGPU::VReg_64RegClass.contains(reg)) {
+          isSGPR = false;
+          width = 2;
+        } else if (AMDGPU::SReg_128RegClass.contains(reg)) {
+          isSGPR = true;
+          width = 4;
+        } else if (AMDGPU::VReg_128RegClass.contains(reg)) {
+          isSGPR = false;
+          width = 4;
+        } else if (AMDGPU::SReg_256RegClass.contains(reg)) {
+          isSGPR = true;
+          width = 8;
+        } else {
+          assert(!"Unknown register class");
+        }
+        hwReg = RI->getEncodingValue(reg);
+        maxUsed = hwReg + width - 1;
+        if (isSGPR) {
+          MaxSGPR = maxUsed > MaxSGPR ? maxUsed : MaxSGPR;
+        } else {
+          MaxVGPR = maxUsed > MaxVGPR ? maxUsed : MaxVGPR;
+        }
+      }
+    }
+  }
+  if (VCCUsed) {
+    MaxSGPR += 2;
+  }
+  SIMachineFunctionInfo * MFI = MF.getInfo<SIMachineFunctionInfo>();
+  OutStreamer.EmitIntValue(MaxSGPR + 1, 4);
+  OutStreamer.EmitIntValue(MaxVGPR + 1, 4);
+  OutStreamer.EmitIntValue(MFI->SPIPSInputAddr, 4);
+}
diff --git a/lib/Target/R600/AMDGPUAsmPrinter.h b/lib/Target/R600/AMDGPUAsmPrinter.h
new file mode 100644
index 0000000000..3812282b17
--- /dev/null
+++ b/lib/Target/R600/AMDGPUAsmPrinter.h
@@ -0,0 +1,44 @@
+//===-- AMDGPUAsmPrinter.h - Print AMDGPU assembly code -------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief AMDGPU Assembly printer class.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef AMDGPU_ASMPRINTER_H
+#define AMDGPU_ASMPRINTER_H
+
+#include "llvm/CodeGen/AsmPrinter.h"
+
+namespace llvm {
+
+class AMDGPUAsmPrinter : public AsmPrinter {
+
+public:
+  explicit AMDGPUAsmPrinter(TargetMachine &TM, MCStreamer &Streamer)
+    : AsmPrinter(TM, Streamer) { }
+
+  virtual bool runOnMachineFunction(MachineFunction &MF);
+
+  virtual const char *getPassName() const {
+    return "AMDGPU Assembly Printer";
+  }
+
+  /// \brief Emit register usage information so that the GPU driver
+  /// can correctly setup the GPU state.
+  void EmitProgramInfo(MachineFunction &MF);
+
+  /// Implemented in AMDGPUMCInstLower.cpp
+  virtual void EmitInstruction(const MachineInstr *MI);
+};
+
+} // End anonymous llvm
+
+#endif //AMDGPU_ASMPRINTER_H
diff --git a/lib/Target/R600/AMDGPUCodeEmitter.h b/lib/Target/R600/AMDGPUCodeEmitter.h
new file mode 100644
index 0000000000..84f3588496
--- /dev/null
+++ b/lib/Target/R600/AMDGPUCodeEmitter.h
@@ -0,0 +1,49 @@
+//===-- AMDGPUCodeEmitter.h - AMDGPU Code Emitter interface -----------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief CodeEmitter interface for R600 and SI codegen.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef AMDGPUCODEEMITTER_H
+#define AMDGPUCODEEMITTER_H
+
+namespace llvm {
+
+class AMDGPUCodeEmitter {
+public:
+  uint64_t getBinaryCodeForInstr(const MachineInstr &MI) const;
+  virtual uint64_t getMachineOpValue(const MachineInstr &MI,
+                                   const MachineOperand &MO) const { return 0; }
+  virtual unsigned GPR4AlignEncode(const MachineInstr  &MI,
+                                     unsigned OpNo) const {
+    return 0;
+  }
+  virtual unsigned GPR2AlignEncode(const MachineInstr &MI,
+                                   unsigned OpNo) const {
+    return 0;
+  }
+  virtual uint64_t VOPPostEncode(const MachineInstr &MI,
+                                 uint64_t Value) const {
+    return Value;
+  }
+  virtual uint64_t i32LiteralEncode(const MachineInstr &MI,
+                                    unsigned OpNo) const {
+    return 0;
+  }
+  virtual uint32_t SMRDmemriEncode(const MachineInstr &MI, unsigned OpNo)
+                                                                   const {
+    return 0;
+  }
+};
+
+} // End namespace llvm
+
+#endif // AMDGPUCODEEMITTER_H
diff --git a/lib/Target/R600/AMDGPUConvertToISA.cpp b/lib/Target/R600/AMDGPUConvertToISA.cpp
new file mode 100644
index 0000000000..50297d1f60
--- /dev/null
+++ b/lib/Target/R600/AMDGPUConvertToISA.cpp
@@ -0,0 +1,62 @@
+//===-- AMDGPUConvertToISA.cpp - Lower AMDIL to HW ISA --------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief This pass lowers AMDIL machine instructions to the appropriate
+/// hardware instructions.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "AMDGPUInstrInfo.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+
+using namespace llvm;
+
+namespace {
+
+class AMDGPUConvertToISAPass : public MachineFunctionPass {
+
+private:
+  static char ID;
+  TargetMachine &TM;
+
+public:
+  AMDGPUConvertToISAPass(TargetMachine &tm) :
+    MachineFunctionPass(ID), TM(tm) { }
+
+  virtual bool runOnMachineFunction(MachineFunction &MF);
+
+  virtual const char *getPassName() const {return "AMDGPU Convert to ISA";}
+
+};
+
+} // End anonymous namespace
+
+char AMDGPUConvertToISAPass::ID = 0;
+
+FunctionPass *llvm::createAMDGPUConvertToISAPass(TargetMachine &tm) {
+  return new AMDGPUConvertToISAPass(tm);
+}
+
+bool AMDGPUConvertToISAPass::runOnMachineFunction(MachineFunction &MF) {
+  const AMDGPUInstrInfo * TII =
+                      static_cast<const AMDGPUInstrInfo*>(TM.getInstrInfo());
+
+  for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
+                                                  BB != BB_E; ++BB) {
+    MachineBasicBlock &MBB = *BB;
+    for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
+                                                      I != E; ++I) {
+      MachineInstr &MI = *I;
+      TII->convertToISA(MI, MF, MBB.findDebugLoc(I));
+    }
+  }
+  return false;
+}
diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
new file mode 100644
index 0000000000..473dac4ddc
--- /dev/null
+++ b/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -0,0 +1,417 @@
+//===-- AMDGPUISelLowering.cpp - AMDGPU Common DAG lowering functions -----===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief This is the parent TargetLowering class for hardware code gen
+/// targets.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUISelLowering.h"
+#include "AMDILIntrinsicInfo.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/SelectionDAG.h"
+#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
+
+using namespace llvm;
+
+AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
+  TargetLowering(TM, new TargetLoweringObjectFileELF()) {
+
+  // Initialize target lowering borrowed from AMDIL
+  InitAMDILLowering();
+
+  // We need to custom lower some of the intrinsics
+  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
+
+  // Library functions.  These default to Expand, but we have instructions
+  // for them.
+  setOperationAction(ISD::FCEIL,  MVT::f32, Legal);
+  setOperationAction(ISD::FEXP2,  MVT::f32, Legal);
+  setOperationAction(ISD::FPOW,   MVT::f32, Legal);
+  setOperationAction(ISD::FLOG2,  MVT::f32, Legal);
+  setOperationAction(ISD::FABS,   MVT::f32, Legal);
+  setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
+  setOperationAction(ISD::FRINT,  MVT::f32, Legal);
+
+  // Lower floating point store/load to integer store/load to reduce the number
+  // of patterns in tablegen.
+  setOperationAction(ISD::STORE, MVT::f32, Promote);
+  AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
+
+  setOperationAction(ISD::STORE, MVT::v4f32, Promote);
+  AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
+
+  setOperationAction(ISD::LOAD, MVT::f32, Promote);
+  AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
+
+  setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
+  AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
+
+  setOperationAction(ISD::UDIV, MVT::i32, Expand);
+  setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
+  setOperationAction(ISD::UREM, MVT::i32, Expand);
+}
+
+//===---------------------------------------------------------------------===//
+// TargetLowering Callbacks
+//===---------------------------------------------------------------------===//
+
+SDValue AMDGPUTargetLowering::LowerFormalArguments(
+                                      SDValue Chain,
+                                      CallingConv::ID CallConv,
+                                      bool isVarArg,
+                                      const SmallVectorImpl<ISD::InputArg> &Ins,
+                                      DebugLoc DL, SelectionDAG &DAG,
+                                      SmallVectorImpl<SDValue> &InVals) const {
+  for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
+    InVals.push_back(SDValue());
+  }
+  return Chain;
+}
+
+SDValue AMDGPUTargetLowering::LowerReturn(
+                                     SDValue Chain,
+                                     CallingConv::ID CallConv,
+                                     bool isVarArg,
+                                     const SmallVectorImpl<ISD::OutputArg> &Outs,
+                                     const SmallVectorImpl<SDValue> &OutVals,
+                                     DebugLoc DL, SelectionDAG &DAG) const {
+  return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
+}
+
+//===---------------------------------------------------------------------===//
+// Target specific lowering
+//===---------------------------------------------------------------------===//
+
+SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
+    const {
+  switch (Op.getOpcode()) {
+  default:
+    Op.getNode()->dump();
+    assert(0 && "Custom lowering code for this"
+        "instruction is not implemented yet!");
+    break;
+  // AMDIL DAG lowering
+  case ISD::SDIV: return LowerSDIV(Op, DAG);
+  case ISD::SREM: return LowerSREM(Op, DAG);
+  case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
+  case ISD::BRCOND: return LowerBRCOND(Op, DAG);
+  // AMDGPU DAG lowering
+  case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
+  case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
+  }
+  return Op;
+}
+
+SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
+    SelectionDAG &DAG) const {
+  unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
+  DebugLoc DL = Op.getDebugLoc();
+  EVT VT = Op.getValueType();
+
+  switch (IntrinsicID) {
+    default: return Op;
+    case AMDGPUIntrinsic::AMDIL_abs:
+      return LowerIntrinsicIABS(Op, DAG);
+    case AMDGPUIntrinsic::AMDIL_exp:
+      return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
+    case AMDGPUIntrinsic::AMDGPU_lrp:
+      return LowerIntrinsicLRP(Op, DAG);
+    case AMDGPUIntrinsic::AMDIL_fraction:
+      return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
+    case AMDGPUIntrinsic::AMDIL_mad:
+      return DAG.getNode(AMDGPUISD::MAD, DL, VT, Op.getOperand(1),
+                              Op.getOperand(2), Op.getOperand(3));
+    case AMDGPUIntrinsic::AMDIL_max:
+      return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
+                                                  Op.getOperand(2));
+    case AMDGPUIntrinsic::AMDGPU_imax:
+      return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
+                                                  Op.getOperand(2));
+    case AMDGPUIntrinsic::AMDGPU_umax:
+      return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
+                                                  Op.getOperand(2));
+    case AMDGPUIntrinsic::AMDIL_min:
+      return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
+                                                  Op.getOperand(2));
+    case AMDGPUIntrinsic::AMDGPU_imin:
+      return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
+                                                  Op.getOperand(2));
+    case AMDGPUIntrinsic::AMDGPU_umin:
+      return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
+                                                  Op.getOperand(2));
+    case AMDGPUIntrinsic::AMDIL_round_nearest:
+      return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
+  }
+}
+
+///IABS(a) = SMAX(sub(0, a), a)
+SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
+    SelectionDAG &DAG) const {
+
+  DebugLoc DL = Op.getDebugLoc();
+  EVT VT = Op.getValueType();
+  SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
+                                              Op.getOperand(1));
+
+  return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
+}
+
+/// Linear Interpolation
+/// LRP(a, b, c) = muladd(a,  b, (1 - a) * c)
+SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
+    SelectionDAG &DAG) const {
+  DebugLoc DL = Op.getDebugLoc();
+  EVT VT = Op.getValueType();
+  SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
+                                DAG.getConstantFP(1.0f, MVT::f32),
+                                Op.getOperand(1));
+  SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
+                                                    Op.getOperand(3));
+  return DAG.getNode(AMDGPUISD::MAD, DL, VT, Op.getOperand(1),
+                                               Op.getOperand(2),
+                                               OneSubAC);
+}
+
+/// \brief Generate Min/Max node
+SDValue AMDGPUTargetLowering::LowerMinMax(SDValue Op,
+    SelectionDAG &DAG) const {
+  DebugLoc DL = Op.getDebugLoc();
+  EVT VT = Op.getValueType();
+
+  SDValue LHS = Op.getOperand(0);
+  SDValue RHS = Op.getOperand(1);
+  SDValue True = Op.getOperand(2);
+  SDValue False = Op.getOperand(3);
+  SDValue CC = Op.getOperand(4);
+
+  if (VT != MVT::f32 ||
+      !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
+    return SDValue();
+  }
+
+  ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
+  switch (CCOpcode) {
+  case ISD::SETOEQ:
+  case ISD::SETONE:
+  case ISD::SETUNE:
+  case ISD::SETNE:
+  case ISD::SETUEQ:
+  case ISD::SETEQ:
+  case ISD::SETFALSE:
+  case ISD::SETFALSE2:
+  case ISD::SETTRUE:
+  case ISD::SETTRUE2:
+  case ISD::SETUO:
+  case ISD::SETO:
+    assert(0 && "Operation should already be optimised !");
+  case ISD::SETULE:
+  case ISD::SETULT:
+  case ISD::SETOLE:
+  case ISD::SETOLT:
+  case ISD::SETLE:
+  case ISD::SETLT: {
+    if (LHS == True)
+      return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
+    else
+      return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
+  }
+  case ISD::SETGT:
+  case ISD::SETGE:
+  case ISD::SETUGE:
+  case ISD::SETOGE:
+  case ISD::SETUGT:
+  case ISD::SETOGT: {
+    if (LHS == True)
+      return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
+    else
+      return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
+  }
+  case ISD::SETCC_INVALID:
+    assert(0 && "Invalid setcc condcode !");
+  }
+  return Op;
+}
+
+
+
+SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
+    SelectionDAG &DAG) const {
+  DebugLoc DL = Op.getDebugLoc();
+  EVT VT = Op.getValueType();
+
+  SDValue Num = Op.getOperand(0);
+  SDValue Den = Op.getOperand(1);
+
+  SmallVector<SDValue, 8> Results;
+
+  // RCP =  URECIP(Den) = 2^32 / Den + e
+  // e is rounding error.
+  SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
+
+  // RCP_LO = umulo(RCP, Den) */
+  SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
+
+  // RCP_HI = mulhu (RCP, Den) */
+  SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
+
+  // NEG_RCP_LO = -RCP_LO
+  SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
+                                                     RCP_LO);
+
+  // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
+  SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
+                                           NEG_RCP_LO, RCP_LO,
+                                           ISD::SETEQ);
+  // Calculate the rounding error from the URECIP instruction
+  // E = mulhu(ABS_RCP_LO, RCP)
+  SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
+
+  // RCP_A_E = RCP + E
+  SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
+
+  // RCP_S_E = RCP - E
+  SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
+
+  // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
+  SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
+                                     RCP_A_E, RCP_S_E,
+                                     ISD::SETEQ);
+  // Quotient = mulhu(Tmp0, Num)
+  SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
+
+  // Num_S_Remainder = Quotient * Den
+  SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
+
+  // Remainder = Num - Num_S_Remainder
+  SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
+
+  // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
+  SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
+                                                 DAG.getConstant(-1, VT),
+                                                 DAG.getConstant(0, VT),
+                                                 ISD::SETGE);
+  // Remainder_GE_Zero = (Remainder >= 0 ? -1 : 0)
+  SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Remainder,
+                                                  DAG.getConstant(0, VT),
+                                                  DAG.getConstant(-1, VT),
+                                                  DAG.getConstant(0, VT),
+                                                  ISD::SETGE);
+  // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
+  SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
+                                               Remainder_GE_Zero);
+
+  // Calculate Division result:
+
+  // Quotient_A_One = Quotient + 1
+  SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
+                                                         DAG.getConstant(1, VT));
+
+  // Quotient_S_One = Quotient - 1
+  SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
+                                                         DAG.getConstant(1, VT));
+
+  // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
+  SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
+                                     Quotient, Quotient_A_One, ISD::SETEQ);
+
+  // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
+  Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
+                            Quotient_S_One, Div, ISD::SETEQ);
+
+  // Calculate Rem result:
+
+  // Remainder_S_Den = Remainder - Den
+  SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
+
+  // Remainder_A_Den = Remainder + Den
+  SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
+
+  // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
+  SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
+                                    Remainder, Remainder_S_Den, ISD::SETEQ);
+
+  // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
+  Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
+                            Remainder_A_Den, Rem, ISD::SETEQ);
+  SDValue Ops[2];
+  Ops[0] = Div;
+  Ops[1] = Rem;
+  return DAG.getMergeValues(Ops, 2, DL);
+}
+
+//===----------------------------------------------------------------------===//
+// Helper functions
+//===----------------------------------------------------------------------===//
+
+bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
+  if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
+    return CFP->isExactlyValue(1.0);
+  }
+  if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
+    return C->isAllOnesValue();
+  }
+  return false;
+}
+
+bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
+  if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
+    return CFP->getValueAPF().isZero();
+  }
+  if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
+    return C->isNullValue();
+  }
+  return false;
+}
+
+SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
+                                                  const TargetRegisterClass *RC,
+                                                   unsigned Reg, EVT VT) const {
+  MachineFunction &MF = DAG.getMachineFunction();
+  MachineRegisterInfo &MRI = MF.getRegInfo();
+  unsigned VirtualRegister;
+  if (!MRI.isLiveIn(Reg)) {
+    VirtualRegister = MRI.createVirtualRegister(RC);
+    MRI.addLiveIn(Reg, VirtualRegister);
+  } else {
+    VirtualRegister = MRI.getLiveInVirtReg(Reg);
+  }
+  return DAG.getRegister(VirtualRegister, VT);
+}
+
+#define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
+
+const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
+  switch (Opcode) {
+  default: return 0;
+  // AMDIL DAG nodes
+  NODE_NAME_CASE(MAD);
+  NODE_NAME_CASE(CALL);
+  NODE_NAME_CASE(UMUL);
+  NODE_NAME_CASE(DIV_INF);
+  NODE_NAME_CASE(RET_FLAG);
+  NODE_NAME_CASE(BRANCH_COND);
+
+  // AMDGPU DAG nodes
+  NODE_NAME_CASE(DWORDADDR)
+  NODE_NAME_CASE(FRACT)
+  NODE_NAME_CASE(FMAX)
+  NODE_NAME_CASE(SMAX)
+  NODE_NAME_CASE(UMAX)
+  NODE_NAME_CASE(FMIN)
+  NODE_NAME_CASE(SMIN)
+  NODE_NAME_CASE(UMIN)
+  NODE_NAME_CASE(URECIP)
+  NODE_NAME_CASE(INTERP)
+  NODE_NAME_CASE(INTERP_P0)
+  NODE_NAME_CASE(EXPORT)
+  }
+}
diff --git a/lib/Target/R600/AMDGPUISelLowering.h b/lib/Target/R600/AMDGPUISelLowering.h
new file mode 100644
index 0000000000..c7abaf69b4
--- /dev/null
+++ b/lib/Target/R600/AMDGPUISelLowering.h
@@ -0,0 +1,144 @@
+//===-- AMDGPUISelLowering.h - AMDGPU Lowering Interface --------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Interface definition of the TargetLowering class that is common
+/// to all AMD GPUs.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef AMDGPUISELLOWERING_H
+#define AMDGPUISELLOWERING_H
+
+#include "llvm/Target/TargetLowering.h"
+
+namespace llvm {
+
+class MachineRegisterInfo;
+
+class AMDGPUTargetLowering : public TargetLowering {
+private:
+  SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerUDIVREM(SDValue Op, SelectionDAG &DAG) const;
+
+protected:
+
+  /// \brief Helper function that adds Reg to the LiveIn list of the DAG's
+  /// MachineFunction.
+  ///
+  /// \returns a RegisterSDNode representing Reg.
+  SDValue CreateLiveInRegister(SelectionDAG &DAG, const TargetRegisterClass *RC,
+                                                  unsigned Reg, EVT VT) const;
+
+  bool isHWTrueValue(SDValue Op) const;
+  bool isHWFalseValue(SDValue Op) const;
+
+public:
+  AMDGPUTargetLowering(TargetMachine &TM);
+
+  virtual SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv,
+                             bool isVarArg,
+                             const SmallVectorImpl<ISD::InputArg> &Ins,
+                             DebugLoc DL, SelectionDAG &DAG,
+                             SmallVectorImpl<SDValue> &InVals) const;
+
+  virtual SDValue LowerReturn(SDValue Chain, CallingConv::ID CallConv,
+                              bool isVarArg,
+                              const SmallVectorImpl<ISD::OutputArg> &Outs,
+                              const SmallVectorImpl<SDValue> &OutVals,
+                              DebugLoc DL, SelectionDAG &DAG) const;
+
+  virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerIntrinsicIABS(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerIntrinsicLRP(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerMinMax(SDValue Op, SelectionDAG &DAG) const;
+  virtual const char* getTargetNodeName(unsigned Opcode) const;
+
+// Functions defined in AMDILISelLowering.cpp
+public:
+
+  /// \brief Determine which of the bits specified in \p Mask are known to be
+  /// either zero or one and return them in the \p KnownZero and \p KnownOne
+  /// bitsets.
+  virtual void computeMaskedBitsForTargetNode(const SDValue Op,
+                                              APInt &KnownZero,
+                                              APInt &KnownOne,
+                                              const SelectionDAG &DAG,
+                                              unsigned Depth = 0) const;
+
+  virtual bool getTgtMemIntrinsic(IntrinsicInfo &Info,
+                                  const CallInst &I, unsigned Intrinsic) const;
+
+  /// We want to mark f32/f64 floating point values as legal.
+  bool isFPImmLegal(const APFloat &Imm, EVT VT) const;
+
+  /// We don't want to shrink f64/f32 constants.
+  bool ShouldShrinkFPConstant(EVT VT) const;
+
+private:
+  void InitAMDILLowering();
+  SDValue LowerSREM(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerSREM8(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerSREM16(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerSREM32(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerSREM64(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerSDIV24(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerSDIV32(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerSDIV64(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
+  EVT genIntType(uint32_t size = 32, uint32_t numEle = 1) const;
+  SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
+};
+
+namespace AMDGPUISD {
+
+enum {
+  // AMDIL ISD Opcodes
+  FIRST_NUMBER = ISD::BUILTIN_OP_END,
+  MAD,         // 32bit Fused Multiply Add instruction
+  CALL,        // Function call based on a single integer
+  UMUL,        // 32bit unsigned multiplication
+  DIV_INF,      // Divide with infinity returned on zero divisor
+  RET_FLAG,
+  BRANCH_COND,
+  // End AMDIL ISD Opcodes
+  BITALIGN,
+  DWORDADDR,
+  FRACT,
+  FMAX,
+  SMAX,
+  UMAX,
+  FMIN,
+  SMIN,
+  UMIN,
+  URECIP,
+  INTERP,
+  INTERP_P0,
+  EXPORT,
+  LAST_AMDGPU_ISD_NUMBER
+};
+
+
+} // End namespace AMDGPUISD
+
+namespace SIISD {
+
+enum {
+  SI_FIRST = AMDGPUISD::LAST_AMDGPU_ISD_NUMBER,
+  VCC_AND,
+  VCC_BITCAST
+};
+
+} // End namespace SIISD
+
+} // End namespace llvm
+
+#endif // AMDGPUISELLOWERING_H
diff --git a/lib/Target/R600/AMDGPUInstrInfo.cpp b/lib/Target/R600/AMDGPUInstrInfo.cpp
new file mode 100644
index 0000000000..e42a46d839
--- /dev/null
+++ b/lib/Target/R600/AMDGPUInstrInfo.cpp
@@ -0,0 +1,257 @@
+//===-- AMDGPUInstrInfo.cpp - Base class for AMD GPU InstrInfo ------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Implementation of the TargetInstrInfo class that is common to all
+/// AMD GPUs.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUInstrInfo.h"
+#include "AMDGPURegisterInfo.h"
+#include "AMDGPUTargetMachine.h"
+#include "AMDIL.h"
+#include "llvm/CodeGen/MachineFrameInfo.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+
+#define GET_INSTRINFO_CTOR
+#include "AMDGPUGenInstrInfo.inc"
+
+using namespace llvm;
+
+AMDGPUInstrInfo::AMDGPUInstrInfo(TargetMachine &tm)
+  : AMDGPUGenInstrInfo(0,0), RI(tm, *this), TM(tm) { }
+
+const AMDGPURegisterInfo &AMDGPUInstrInfo::getRegisterInfo() const {
+  return RI;
+}
+
+bool AMDGPUInstrInfo::isCoalescableExtInstr(const MachineInstr &MI,
+                                           unsigned &SrcReg, unsigned &DstReg,
+                                           unsigned &SubIdx) const {
+// TODO: Implement this function
+  return false;
+}
+
+unsigned AMDGPUInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
+                                             int &FrameIndex) const {
+// TODO: Implement this function
+  return 0;
+}
+
+unsigned AMDGPUInstrInfo::isLoadFromStackSlotPostFE(const MachineInstr *MI,
+                                                   int &FrameIndex) const {
+// TODO: Implement this function
+  return 0;
+}
+
+bool AMDGPUInstrInfo::hasLoadFromStackSlot(const MachineInstr *MI,
+                                          const MachineMemOperand *&MMO,
+                                          int &FrameIndex) const {
+// TODO: Implement this function
+  return false;
+}
+unsigned AMDGPUInstrInfo::isStoreFromStackSlot(const MachineInstr *MI,
+                                              int &FrameIndex) const {
+// TODO: Implement this function
+  return 0;
+}
+unsigned AMDGPUInstrInfo::isStoreFromStackSlotPostFE(const MachineInstr *MI,
+                                                    int &FrameIndex) const {
+// TODO: Implement this function
+  return 0;
+}
+bool AMDGPUInstrInfo::hasStoreFromStackSlot(const MachineInstr *MI,
+                                           const MachineMemOperand *&MMO,
+                                           int &FrameIndex) const {
+// TODO: Implement this function
+  return false;
+}
+
+MachineInstr *
+AMDGPUInstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI,
+                                      MachineBasicBlock::iterator &MBBI,
+                                      LiveVariables *LV) const {
+// TODO: Implement this function
+  return NULL;
+}
+bool AMDGPUInstrInfo::getNextBranchInstr(MachineBasicBlock::iterator &iter,
+                                        MachineBasicBlock &MBB) const {
+  while (iter != MBB.end()) {
+    switch (iter->getOpcode()) {
+    default:
+      break;
+    case AMDGPU::BRANCH_COND_i32:
+    case AMDGPU::BRANCH_COND_f32:
+    case AMDGPU::BRANCH:
+      return true;
+    };
+    ++iter;
+  }
+  return false;
+}
+
+MachineBasicBlock::iterator skipFlowControl(MachineBasicBlock *MBB) {
+  MachineBasicBlock::iterator tmp = MBB->end();
+  if (!MBB->size()) {
+    return MBB->end();
+  }
+  while (--tmp) {
+    if (tmp->getOpcode() == AMDGPU::ENDLOOP
+        || tmp->getOpcode() == AMDGPU::ENDIF
+        || tmp->getOpcode() == AMDGPU::ELSE) {
+      if (tmp == MBB->begin()) {
+        return tmp;
+      } else {
+        continue;
+      }
+    }  else {
+      return ++tmp;
+    }
+  }
+  return MBB->end();
+}
+
+void
+AMDGPUInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
+                                    MachineBasicBlock::iterator MI,
+                                    unsigned SrcReg, bool isKill,
+                                    int FrameIndex,
+                                    const TargetRegisterClass *RC,
+                                    const TargetRegisterInfo *TRI) const {
+  assert(!"Not Implemented");
+}
+
+void
+AMDGPUInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
+                                     MachineBasicBlock::iterator MI,
+                                     unsigned DestReg, int FrameIndex,
+                                     const TargetRegisterClass *RC,
+                                     const TargetRegisterInfo *TRI) const {
+  assert(!"Not Implemented");
+}
+
+MachineInstr *
+AMDGPUInstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
+                                      MachineInstr *MI,
+                                      const SmallVectorImpl<unsigned> &Ops,
+                                      int FrameIndex) const {
+// TODO: Implement this function
+  return 0;
+}
+MachineInstr*
+AMDGPUInstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
+                                      MachineInstr *MI,
+                                      const SmallVectorImpl<unsigned> &Ops,
+                                      MachineInstr *LoadMI) const {
+  // TODO: Implement this function
+  return 0;
+}
+bool
+AMDGPUInstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
+                                     const SmallVectorImpl<unsigned> &Ops) const {
+  // TODO: Implement this function
+  return false;
+}
+bool
+AMDGPUInstrInfo::unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
+                                 unsigned Reg, bool UnfoldLoad,
+                                 bool UnfoldStore,
+                                 SmallVectorImpl<MachineInstr*> &NewMIs) const {
+  // TODO: Implement this function
+  return false;
+}
+
+bool
+AMDGPUInstrInfo::unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
+                                    SmallVectorImpl<SDNode*> &NewNodes) const {
+  // TODO: Implement this function
+  return false;
+}
+
+unsigned
+AMDGPUInstrInfo::getOpcodeAfterMemoryUnfold(unsigned Opc,
+                                           bool UnfoldLoad, bool UnfoldStore,
+                                           unsigned *LoadRegIndex) const {
+  // TODO: Implement this function
+  return 0;
+}
+
+bool AMDGPUInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
+                                             int64_t Offset1, int64_t Offset2,
+                                             unsigned NumLoads) const {
+  assert(Offset2 > Offset1
+         && "Second offset should be larger than first offset!");
+  // If we have less than 16 loads in a row, and the offsets are within 16,
+  // then schedule together.
+  // TODO: Make the loads schedule near if it fits in a cacheline
+  return (NumLoads < 16 && (Offset2 - Offset1) < 16);
+}
+
+bool
+AMDGPUInstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond)
+  const {
+  // TODO: Implement this function
+  return true;
+}
+void AMDGPUInstrInfo::insertNoop(MachineBasicBlock &MBB,
+                                MachineBasicBlock::iterator MI) const {
+  // TODO: Implement this function
+}
+
+bool AMDGPUInstrInfo::isPredicated(const MachineInstr *MI) const {
+  // TODO: Implement this function
+  return false;
+}
+bool
+AMDGPUInstrInfo::SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
+                                  const SmallVectorImpl<MachineOperand> &Pred2)
+  const {
+  // TODO: Implement this function
+  return false;
+}
+
+bool AMDGPUInstrInfo::DefinesPredicate(MachineInstr *MI,
+                                      std::vector<MachineOperand> &Pred) const {
+  // TODO: Implement this function
+  return false;
+}
+
+bool AMDGPUInstrInfo::isPredicable(MachineInstr *MI) const {
+  // TODO: Implement this function
+  return MI->getDesc().isPredicable();
+}
+
+bool
+AMDGPUInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
+  // TODO: Implement this function
+  return true;
+}
+ 
+void AMDGPUInstrInfo::convertToISA(MachineInstr & MI, MachineFunction &MF,
+    DebugLoc DL) const {
+  MachineRegisterInfo &MRI = MF.getRegInfo();
+  const AMDGPURegisterInfo & RI = getRegisterInfo();
+
+  for (unsigned i = 0; i < MI.getNumOperands(); i++) {
+    MachineOperand &MO = MI.getOperand(i);
+    // Convert dst regclass to one that is supported by the ISA
+    if (MO.isReg() && MO.isDef()) {
+      if (TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
+        const TargetRegisterClass * oldRegClass = MRI.getRegClass(MO.getReg());
+        const TargetRegisterClass * newRegClass = RI.getISARegClass(oldRegClass);
+
+        assert(newRegClass);
+
+        MRI.setRegClass(MO.getReg(), newRegClass);
+      }
+    }
+  }
+}
diff --git a/lib/Target/R600/AMDGPUInstrInfo.h b/lib/Target/R600/AMDGPUInstrInfo.h
new file mode 100644
index 0000000000..32ac691fe0
--- /dev/null
+++ b/lib/Target/R600/AMDGPUInstrInfo.h
@@ -0,0 +1,149 @@
+//===-- AMDGPUInstrInfo.h - AMDGPU Instruction Information ------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Contains the definition of a TargetInstrInfo class that is common
+/// to all AMD GPUs.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef AMDGPUINSTRUCTIONINFO_H
+#define AMDGPUINSTRUCTIONINFO_H
+
+#include "AMDGPURegisterInfo.h"
+#include "AMDGPUInstrInfo.h"
+#include "llvm/Target/TargetInstrInfo.h"
+
+#include <map>
+
+#define GET_INSTRINFO_HEADER
+#define GET_INSTRINFO_ENUM
+#include "AMDGPUGenInstrInfo.inc"
+
+#define OPCODE_IS_ZERO_INT AMDGPU::PRED_SETE_INT
+#define OPCODE_IS_NOT_ZERO_INT AMDGPU::PRED_SETNE_INT
+#define OPCODE_IS_ZERO AMDGPU::PRED_SETE
+#define OPCODE_IS_NOT_ZERO AMDGPU::PRED_SETNE
+
+namespace llvm {
+
+class AMDGPUTargetMachine;
+class MachineFunction;
+class MachineInstr;
+class MachineInstrBuilder;
+
+class AMDGPUInstrInfo : public AMDGPUGenInstrInfo {
+private:
+  const AMDGPURegisterInfo RI;
+  TargetMachine &TM;
+  bool getNextBranchInstr(MachineBasicBlock::iterator &iter,
+                          MachineBasicBlock &MBB) const;
+public:
+  explicit AMDGPUInstrInfo(TargetMachine &tm);
+
+  virtual const AMDGPURegisterInfo &getRegisterInfo() const = 0;
+
+  bool isCoalescableExtInstr(const MachineInstr &MI, unsigned &SrcReg,
+                             unsigned &DstReg, unsigned &SubIdx) const;
+
+  unsigned isLoadFromStackSlot(const MachineInstr *MI, int &FrameIndex) const;
+  unsigned isLoadFromStackSlotPostFE(const MachineInstr *MI,
+                                     int &FrameIndex) const;
+  bool hasLoadFromStackSlot(const MachineInstr *MI,
+                            const MachineMemOperand *&MMO,
+                            int &FrameIndex) const;
+  unsigned isStoreFromStackSlot(const MachineInstr *MI, int &FrameIndex) const;
+  unsigned isStoreFromStackSlotPostFE(const MachineInstr *MI,
+                                      int &FrameIndex) const;
+  bool hasStoreFromStackSlot(const MachineInstr *MI,
+                             const MachineMemOperand *&MMO,
+                             int &FrameIndex) const;
+
+  MachineInstr *
+  convertToThreeAddress(MachineFunction::iterator &MFI,
+                        MachineBasicBlock::iterator &MBBI,
+                        LiveVariables *LV) const;
+
+
+  virtual void copyPhysReg(MachineBasicBlock &MBB,
+                           MachineBasicBlock::iterator MI, DebugLoc DL,
+                           unsigned DestReg, unsigned SrcReg,
+                           bool KillSrc) const = 0;
+
+  void storeRegToStackSlot(MachineBasicBlock &MBB,
+                           MachineBasicBlock::iterator MI,
+                           unsigned SrcReg, bool isKill, int FrameIndex,
+                           const TargetRegisterClass *RC,
+                           const TargetRegisterInfo *TRI) const;
+  void loadRegFromStackSlot(MachineBasicBlock &MBB,
+                            MachineBasicBlock::iterator MI,
+                            unsigned DestReg, int FrameIndex,
+                            const TargetRegisterClass *RC,
+                            const TargetRegisterInfo *TRI) const;
+
+protected:
+  MachineInstr *foldMemoryOperandImpl(MachineFunction &MF,
+                                      MachineInstr *MI,
+                                      const SmallVectorImpl<unsigned> &Ops,
+                                      int FrameIndex) const;
+  MachineInstr *foldMemoryOperandImpl(MachineFunction &MF,
+                                      MachineInstr *MI,
+                                      const SmallVectorImpl<unsigned> &Ops,
+                                      MachineInstr *LoadMI) const;
+public:
+  bool canFoldMemoryOperand(const MachineInstr *MI,
+                            const SmallVectorImpl<unsigned> &Ops) const;
+  bool unfoldMemoryOperand(MachineFunction &MF, MachineInstr *MI,
+                           unsigned Reg, bool UnfoldLoad, bool UnfoldStore,
+                           SmallVectorImpl<MachineInstr *> &NewMIs) const;
+  bool unfoldMemoryOperand(SelectionDAG &DAG, SDNode *N,
+                           SmallVectorImpl<SDNode *> &NewNodes) const;
+  unsigned getOpcodeAfterMemoryUnfold(unsigned Opc,
+                                      bool UnfoldLoad, bool UnfoldStore,
+                                      unsigned *LoadRegIndex = 0) const;
+  bool shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2,
+                               int64_t Offset1, int64_t Offset2,
+                               unsigned NumLoads) const;
+
+  bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const;
+  void insertNoop(MachineBasicBlock &MBB,
+                  MachineBasicBlock::iterator MI) const;
+  bool isPredicated(const MachineInstr *MI) const;
+  bool SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
+                         const SmallVectorImpl<MachineOperand> &Pred2) const;
+  bool DefinesPredicate(MachineInstr *MI,
+                        std::vector<MachineOperand> &Pred) const;
+  bool isPredicable(MachineInstr *MI) const;
+  bool isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const;
+
+  // Helper functions that check the opcode for status information
+  bool isLoadInst(llvm::MachineInstr *MI) const;
+  bool isExtLoadInst(llvm::MachineInstr *MI) const;
+  bool isSWSExtLoadInst(llvm::MachineInstr *MI) const;
+  bool isSExtLoadInst(llvm::MachineInstr *MI) const;
+  bool isZExtLoadInst(llvm::MachineInstr *MI) const;
+  bool isAExtLoadInst(llvm::MachineInstr *MI) const;
+  bool isStoreInst(llvm::MachineInstr *MI) const;
+  bool isTruncStoreInst(llvm::MachineInstr *MI) const;
+
+  virtual MachineInstr* getMovImmInstr(MachineFunction *MF, unsigned DstReg,
+                                       int64_t Imm) const = 0;
+  virtual unsigned getIEQOpcode() const = 0;
+  virtual bool isMov(unsigned opcode) const = 0;
+
+  /// \brief Convert the AMDIL MachineInstr to a supported ISA
+  /// MachineInstr
+  virtual void convertToISA(MachineInstr & MI, MachineFunction &MF,
+    DebugLoc DL) const;
+
+};
+
+} // End llvm namespace
+
+#endif // AMDGPUINSTRINFO_H
diff --git a/lib/Target/R600/AMDGPUInstrInfo.td b/lib/Target/R600/AMDGPUInstrInfo.td
new file mode 100644
index 0000000000..96368e8541
--- /dev/null
+++ b/lib/Target/R600/AMDGPUInstrInfo.td
@@ -0,0 +1,74 @@
+//===-- AMDGPUInstrInfo.td - AMDGPU DAG nodes --------------*- tablegen -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file contains DAG node defintions for the AMDGPU target.
+//
+//===----------------------------------------------------------------------===//
+
+//===----------------------------------------------------------------------===//
+// AMDGPU DAG Profiles
+//===----------------------------------------------------------------------===//
+
+def AMDGPUDTIntTernaryOp : SDTypeProfile<1, 3, [
+  SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>, SDTCisInt<0>, SDTCisInt<3>
+]>;
+
+//===----------------------------------------------------------------------===//
+// AMDGPU DAG Nodes
+//
+
+// out = ((a << 32) | b) >> c)
+//
+// Can be used to optimize rtol:
+// rotl(a, b) = bitalign(a, a, 32 - b)
+def AMDGPUbitalign : SDNode<"AMDGPUISD::BITALIGN", AMDGPUDTIntTernaryOp>;
+
+// This argument to this node is a dword address.
+def AMDGPUdwordaddr : SDNode<"AMDGPUISD::DWORDADDR", SDTIntUnaryOp>;
+
+// out = a - floor(a)
+def AMDGPUfract : SDNode<"AMDGPUISD::FRACT", SDTFPUnaryOp>;
+
+// out = max(a, b) a and b are floats
+def AMDGPUfmax : SDNode<"AMDGPUISD::FMAX", SDTFPBinOp,
+  [SDNPCommutative, SDNPAssociative]
+>;
+
+// out = max(a, b) a and b are signed ints
+def AMDGPUsmax : SDNode<"AMDGPUISD::SMAX", SDTIntBinOp,
+  [SDNPCommutative, SDNPAssociative]
+>;
+
+// out = max(a, b) a and b are unsigned ints
+def AMDGPUumax : SDNode<"AMDGPUISD::UMAX", SDTIntBinOp,
+  [SDNPCommutative, SDNPAssociative]
+>;
+
+// out = min(a, b) a and b are floats
+def AMDGPUfmin : SDNode<"AMDGPUISD::FMIN", SDTFPBinOp,
+  [SDNPCommutative, SDNPAssociative]
+>;
+
+// out = min(a, b) a snd b are signed ints
+def AMDGPUsmin : SDNode<"AMDGPUISD::SMIN", SDTIntBinOp,
+  [SDNPCommutative, SDNPAssociative]
+>;
+
+// out = min(a, b) a and b are unsigned ints
+def AMDGPUumin : SDNode<"AMDGPUISD::UMIN", SDTIntBinOp,
+  [SDNPCommutative, SDNPAssociative]
+>;
+
+// urecip - This operation is a helper for integer division, it returns the
+// result of 1 / a as a fractional unsigned integer.
+// out = (2^32 / a) + e
+// e is rounding error
+def AMDGPUurecip : SDNode<"AMDGPUISD::URECIP", SDTIntUnaryOp>;
+
+def fpow : SDNode<"ISD::FPOW", SDTFPBinOp>;
diff --git a/lib/Target/R600/AMDGPUInstructions.td b/lib/Target/R600/AMDGPUInstructions.td
new file mode 100644
index 0000000000..e634d20b61
--- /dev/null
+++ b/lib/Target/R600/AMDGPUInstructions.td
@@ -0,0 +1,190 @@
+//===-- AMDGPUInstructions.td - Common instruction defs ---*- tablegen -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file contains instruction defs that are common to all hw codegen
+// targets.
+//
+//===----------------------------------------------------------------------===//
+
+class AMDGPUInst <dag outs, dag ins, string asm, list<dag> pattern> : Instruction {
+  field bits<16> AMDILOp = 0;
+  field bits<3> Gen = 0;
+
+  let Namespace = "AMDGPU";
+  let OutOperandList = outs;
+  let InOperandList = ins;
+  let AsmString = asm;
+  let Pattern = pattern;
+  let Itinerary = NullALU;
+  let TSFlags{42-40} = Gen;
+  let TSFlags{63-48} = AMDILOp;
+}
+
+class AMDGPUShaderInst <dag outs, dag ins, string asm, list<dag> pattern>
+    : AMDGPUInst<outs, ins, asm, pattern> {
+
+  field bits<32> Inst = 0xffffffff;
+
+}
+
+def InstFlag : OperandWithDefaultOps <i32, (ops (i32 0))>;
+
+def COND_EQ : PatLeaf <
+  (cond),
+  [{switch(N->get()){{default: return false;
+                     case ISD::SETOEQ: case ISD::SETUEQ:
+                     case ISD::SETEQ: return true;}}}]
+>;
+
+def COND_NE : PatLeaf <
+  (cond),
+  [{switch(N->get()){{default: return false;
+                     case ISD::SETONE: case ISD::SETUNE:
+                     case ISD::SETNE: return true;}}}]
+>;
+def COND_GT : PatLeaf <
+  (cond),
+  [{switch(N->get()){{default: return false;
+                     case ISD::SETOGT: case ISD::SETUGT:
+                     case ISD::SETGT: return true;}}}]
+>;
+
+def COND_GE : PatLeaf <
+  (cond),
+  [{switch(N->get()){{default: return false;
+                     case ISD::SETOGE: case ISD::SETUGE:
+                     case ISD::SETGE: return true;}}}]
+>;
+
+def COND_LT : PatLeaf <
+  (cond),
+  [{switch(N->get()){{default: return false;
+                     case ISD::SETOLT: case ISD::SETULT:
+                     case ISD::SETLT: return true;}}}]
+>;
+
+def COND_LE : PatLeaf <
+  (cond),
+  [{switch(N->get()){{default: return false;
+                     case ISD::SETOLE: case ISD::SETULE:
+                     case ISD::SETLE: return true;}}}]
+>;
+
+//===----------------------------------------------------------------------===//
+// Load/Store Pattern Fragments
+//===----------------------------------------------------------------------===//
+
+def zextloadi8_global : PatFrag<(ops node:$ptr), (zextloadi8 node:$ptr), [{
+    return isGlobalLoad(dyn_cast<LoadSDNode>(N));
+}]>;
+
+class Constants {
+int TWO_PI = 0x40c90fdb;
+int PI = 0x40490fdb;
+int TWO_PI_INV = 0x3e22f983;
+}
+def CONST : Constants;
+
+def FP_ZERO : PatLeaf <
+  (fpimm),
+  [{return N->getValueAPF().isZero();}]
+>;
+
+def FP_ONE : PatLeaf <
+  (fpimm),
+  [{return N->isExactlyValue(1.0);}]
+>;
+
+let isCodeGenOnly = 1, isPseudo = 1, usesCustomInserter = 1  in {
+
+class CLAMP <RegisterClass rc> : AMDGPUShaderInst <
+  (outs rc:$dst),
+  (ins rc:$src0),
+  "CLAMP $dst, $src0",
+  [(set rc:$dst, (int_AMDIL_clamp rc:$src0, (f32 FP_ZERO), (f32 FP_ONE)))]
+>;
+
+class FABS <RegisterClass rc> : AMDGPUShaderInst <
+  (outs rc:$dst),
+  (ins rc:$src0),
+  "FABS $dst, $src0",
+  [(set rc:$dst, (fabs rc:$src0))]
+>;
+
+class FNEG <RegisterClass rc> : AMDGPUShaderInst <
+  (outs rc:$dst),
+  (ins rc:$src0),
+  "FNEG $dst, $src0",
+  [(set rc:$dst, (fneg rc:$src0))]
+>;
+
+def SHADER_TYPE : AMDGPUShaderInst <
+  (outs),
+  (ins i32imm:$type),
+  "SHADER_TYPE $type",
+  [(int_AMDGPU_shader_type imm:$type)]
+>;
+
+} // End isCodeGenOnly = 1, isPseudo = 1, hasCustomInserter = 1
+
+/* Generic helper patterns for intrinsics */
+/* -------------------------------------- */
+
+class POW_Common <AMDGPUInst log_ieee, AMDGPUInst exp_ieee, AMDGPUInst mul,
+                  RegisterClass rc> : Pat <
+  (fpow rc:$src0, rc:$src1),
+  (exp_ieee (mul rc:$src1, (log_ieee rc:$src0)))
+>;
+
+/* Other helper patterns */
+/* --------------------- */
+
+/* Extract element pattern */
+class Extract_Element <ValueType sub_type, ValueType vec_type,
+                     RegisterClass vec_class, int sub_idx, 
+                     SubRegIndex sub_reg>: Pat<
+  (sub_type (vector_extract (vec_type vec_class:$src), sub_idx)),
+  (EXTRACT_SUBREG vec_class:$src, sub_reg)
+>;
+
+/* Insert element pattern */
+class Insert_Element <ValueType elem_type, ValueType vec_type,
+                      RegisterClass elem_class, RegisterClass vec_class,
+                      int sub_idx, SubRegIndex sub_reg> : Pat <
+
+  (vec_type (vector_insert (vec_type vec_class:$vec),
+                           (elem_type elem_class:$elem), sub_idx)),
+  (INSERT_SUBREG vec_class:$vec, elem_class:$elem, sub_reg)
+>;
+
+// Vector Build pattern
+class Vector_Build <ValueType vecType, RegisterClass vectorClass,
+                    ValueType elemType, RegisterClass elemClass> : Pat <
+  (vecType (build_vector (elemType elemClass:$x), (elemType elemClass:$y),
+                         (elemType elemClass:$z), (elemType elemClass:$w))),
+  (INSERT_SUBREG (INSERT_SUBREG (INSERT_SUBREG (INSERT_SUBREG
+  (vecType (IMPLICIT_DEF)), elemClass:$x, sel_x), elemClass:$y, sel_y),
+                            elemClass:$z, sel_z), elemClass:$w, sel_w)
+>;
+
+// bitconvert pattern
+class BitConvert <ValueType dt, ValueType st, RegisterClass rc> : Pat <
+  (dt (bitconvert (st rc:$src0))),
+  (dt rc:$src0)
+>;
+
+class DwordAddrPat<ValueType vt, RegisterClass rc> : Pat <
+  (vt (AMDGPUdwordaddr (vt rc:$addr))),
+  (vt rc:$addr)
+>;
+
+include "R600Instructions.td"
+
+include "SIInstrInfo.td"
+
diff --git a/lib/Target/R600/AMDGPUIntrinsics.td b/lib/Target/R600/AMDGPUIntrinsics.td
new file mode 100644
index 0000000000..2ba2d4b90d
--- /dev/null
+++ b/lib/Target/R600/AMDGPUIntrinsics.td
@@ -0,0 +1,62 @@
+//===-- AMDGPUIntrinsics.td - Common intrinsics  -*- tablegen -*-----------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines intrinsics that are used by all hw codegen targets.
+//
+//===----------------------------------------------------------------------===//
+
+let TargetPrefix = "AMDGPU", isTarget = 1 in {
+
+  def int_AMDGPU_load_const : Intrinsic<[llvm_float_ty], [llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_load_imm : Intrinsic<[llvm_v4f32_ty], [llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_reserve_reg : Intrinsic<[], [llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_store_output : Intrinsic<[], [llvm_float_ty, llvm_i32_ty], []>;
+  def int_AMDGPU_swizzle : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_i32_ty], [IntrNoMem]>;
+
+  def int_AMDGPU_arl : Intrinsic<[llvm_i32_ty], [llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_cndlt : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_div : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_dp4 : Intrinsic<[llvm_float_ty], [llvm_v4f32_ty, llvm_v4f32_ty], [IntrNoMem]>;
+  def int_AMDGPU_kill : Intrinsic<[], [llvm_float_ty], []>;
+  def int_AMDGPU_kilp : Intrinsic<[], [], []>;
+  def int_AMDGPU_lrp : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_mul : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_pow : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_rcp : Intrinsic<[llvm_float_ty], [llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_rsq : Intrinsic<[llvm_float_ty], [llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_seq : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_sgt : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_sge : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_sle : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_sne : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_mullit : Intrinsic<[llvm_v4f32_ty], [llvm_float_ty, llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_tex : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_txb : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_txf : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_txq : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_txd : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_v4f32_ty, llvm_v4f32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_txl : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_trunc : Intrinsic<[llvm_float_ty], [llvm_float_ty], [IntrNoMem]>;
+  def int_AMDGPU_ddx : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_ddy : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_imax : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_imin : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_umax : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_umin : Intrinsic<[llvm_i32_ty], [llvm_i32_ty, llvm_i32_ty], [IntrNoMem]>;
+  def int_AMDGPU_cube : Intrinsic<[llvm_v4f32_ty], [llvm_v4f32_ty], [IntrNoMem]>;
+
+  def int_AMDGPU_shader_type : Intrinsic<[], [llvm_i32_ty], []>;
+}
+
+let TargetPrefix = "TGSI", isTarget = 1 in {
+
+  def int_TGSI_lit_z : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty, llvm_float_ty],[IntrNoMem]>;
+}
+
+include "SIIntrinsics.td"
diff --git a/lib/Target/R600/AMDGPUMCInstLower.cpp b/lib/Target/R600/AMDGPUMCInstLower.cpp
new file mode 100644
index 0000000000..de4053ea09
--- /dev/null
+++ b/lib/Target/R600/AMDGPUMCInstLower.cpp
@@ -0,0 +1,77 @@
+//===- AMDGPUMCInstLower.cpp - Lower AMDGPU MachineInstr to an MCInst -----===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Code to lower AMDGPU MachineInstrs to their corresponding MCInst.
+//
+//===----------------------------------------------------------------------===//
+//
+
+#include "AMDGPUMCInstLower.h"
+#include "AMDGPUAsmPrinter.h"
+#include "R600InstrInfo.h"
+#include "llvm/CodeGen/MachineBasicBlock.h"
+#include "llvm/CodeGen/MachineInstr.h"
+#include "llvm/Constants.h"
+#include "llvm/MC/MCInst.h"
+#include "llvm/MC/MCStreamer.h"
+#include "llvm/Support/ErrorHandling.h"
+
+using namespace llvm;
+
+AMDGPUMCInstLower::AMDGPUMCInstLower() { }
+
+void AMDGPUMCInstLower::lower(const MachineInstr *MI, MCInst &OutMI) const {
+  OutMI.setOpcode(MI->getOpcode());
+
+  for (unsigned i = 0, e = MI->getNumExplicitOperands(); i != e; ++i) {
+    const MachineOperand &MO = MI->getOperand(i);
+
+    MCOperand MCOp;
+    switch (MO.getType()) {
+    default:
+      llvm_unreachable("unknown operand type");
+    case MachineOperand::MO_FPImmediate: {
+      const APFloat &FloatValue = MO.getFPImm()->getValueAPF();
+      assert(&FloatValue.getSemantics() == &APFloat::IEEEsingle &&
+             "Only floating point immediates are supported at the moment.");
+      MCOp = MCOperand::CreateFPImm(FloatValue.convertToFloat());
+      break;
+    }
+    case MachineOperand::MO_Immediate:
+      MCOp = MCOperand::CreateImm(MO.getImm());
+      break;
+    case MachineOperand::MO_Register:
+      MCOp = MCOperand::CreateReg(MO.getReg());
+      break;
+    }
+    OutMI.addOperand(MCOp);
+  }
+}
+
+void AMDGPUAsmPrinter::EmitInstruction(const MachineInstr *MI) {
+  AMDGPUMCInstLower MCInstLowering;
+
+  if (MI->isBundle()) {
+    const MachineBasicBlock *MBB = MI->getParent();
+    MachineBasicBlock::const_instr_iterator I = MI;
+    ++I;
+    while (I != MBB->end() && I->isInsideBundle()) {
+      MCInst MCBundleInst;
+      const MachineInstr *BundledInst = I;
+      MCInstLowering.lower(BundledInst, MCBundleInst);
+      OutStreamer.EmitInstruction(MCBundleInst);
+      ++I;
+    }
+  } else {
+    MCInst TmpInst;
+    MCInstLowering.lower(MI, TmpInst);
+    OutStreamer.EmitInstruction(TmpInst);
+  }
+}
diff --git a/lib/Target/R600/AMDGPUMCInstLower.h b/lib/Target/R600/AMDGPUMCInstLower.h
new file mode 100644
index 0000000000..d7bf827ff2
--- /dev/null
+++ b/lib/Target/R600/AMDGPUMCInstLower.h
@@ -0,0 +1,31 @@
+//===- AMDGPUMCInstLower.h MachineInstr Lowering Interface ------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//===----------------------------------------------------------------------===//
+
+#ifndef AMDGPU_MCINSTLOWER_H
+#define AMDGPU_MCINSTLOWER_H
+
+namespace llvm {
+
+class MCInst;
+class MachineInstr;
+
+class AMDGPUMCInstLower {
+
+public:
+  AMDGPUMCInstLower();
+
+  /// \brief Lower a MachineInstr to an MCInst
+  void lower(const MachineInstr *MI, MCInst &OutMI) const;
+
+};
+
+} // End namespace llvm
+
+#endif //AMDGPU_MCINSTLOWER_H
diff --git a/lib/Target/R600/AMDGPURegisterInfo.cpp b/lib/Target/R600/AMDGPURegisterInfo.cpp
new file mode 100644
index 0000000000..eeafec898d
--- /dev/null
+++ b/lib/Target/R600/AMDGPURegisterInfo.cpp
@@ -0,0 +1,51 @@
+//===-- AMDGPURegisterInfo.cpp - AMDGPU Register Information -------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Parent TargetRegisterInfo class common to all hw codegen targets.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPURegisterInfo.h"
+#include "AMDGPUTargetMachine.h"
+
+using namespace llvm;
+
+AMDGPURegisterInfo::AMDGPURegisterInfo(TargetMachine &tm,
+    const TargetInstrInfo &tii)
+: AMDGPUGenRegisterInfo(0),
+  TM(tm),
+  TII(tii)
+  { }
+
+//===----------------------------------------------------------------------===//
+// Function handling callbacks - Functions are a seldom used feature of GPUS, so
+// they are not supported at this time.
+//===----------------------------------------------------------------------===//
+
+const uint16_t AMDGPURegisterInfo::CalleeSavedReg = AMDGPU::NoRegister;
+
+const uint16_t* AMDGPURegisterInfo::getCalleeSavedRegs(const MachineFunction *MF)
+                                                                         const {
+  return &CalleeSavedReg;
+}
+
+void AMDGPURegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator MI,
+                                             int SPAdj,
+                                             RegScavenger *RS) const {
+  assert(!"Subroutines not supported yet");
+}
+
+unsigned AMDGPURegisterInfo::getFrameRegister(const MachineFunction &MF) const {
+  assert(!"Subroutines not supported yet");
+  return 0;
+}
+
+#define GET_REGINFO_TARGET_DESC
+#include "AMDGPUGenRegisterInfo.inc"
diff --git a/lib/Target/R600/AMDGPURegisterInfo.h b/lib/Target/R600/AMDGPURegisterInfo.h
new file mode 100644
index 0000000000..76ee7ae06a
--- /dev/null
+++ b/lib/Target/R600/AMDGPURegisterInfo.h
@@ -0,0 +1,63 @@
+//===-- AMDGPURegisterInfo.h - AMDGPURegisterInfo Interface -*- C++ -*-----===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief TargetRegisterInfo interface that is implemented by all hw codegen
+/// targets.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef AMDGPUREGISTERINFO_H
+#define AMDGPUREGISTERINFO_H
+
+#include "llvm/ADT/BitVector.h"
+#include "llvm/Target/TargetRegisterInfo.h"
+
+#define GET_REGINFO_HEADER
+#define GET_REGINFO_ENUM
+#include "AMDGPUGenRegisterInfo.inc"
+
+namespace llvm {
+
+class AMDGPUTargetMachine;
+class TargetInstrInfo;
+
+struct AMDGPURegisterInfo : public AMDGPUGenRegisterInfo {
+  TargetMachine &TM;
+  const TargetInstrInfo &TII;
+  static const uint16_t CalleeSavedReg;
+
+  AMDGPURegisterInfo(TargetMachine &tm, const TargetInstrInfo &tii);
+
+  virtual BitVector getReservedRegs(const MachineFunction &MF) const {
+    assert(!"Unimplemented");  return BitVector();
+  }
+
+  /// \param RC is an AMDIL reg class.
+  ///
+  /// \returns The ISA reg class that is equivalent to \p RC.
+  virtual const TargetRegisterClass * getISARegClass(
+                                         const TargetRegisterClass * RC) const {
+    assert(!"Unimplemented"); return NULL;
+  }
+
+  virtual const TargetRegisterClass* getCFGStructurizerRegClass(MVT VT) const {
+    assert(!"Unimplemented"); return NULL;
+  }
+
+  const uint16_t* getCalleeSavedRegs(const MachineFunction *MF) const;
+  void eliminateFrameIndex(MachineBasicBlock::iterator MI, int SPAdj,
+                           RegScavenger *RS) const;
+  unsigned getFrameRegister(const MachineFunction &MF) const;
+
+};
+
+} // End namespace llvm
+
+#endif // AMDIDSAREGISTERINFO_H
diff --git a/lib/Target/R600/AMDGPURegisterInfo.td b/lib/Target/R600/AMDGPURegisterInfo.td
new file mode 100644
index 0000000000..8181e023aa
--- /dev/null
+++ b/lib/Target/R600/AMDGPURegisterInfo.td
@@ -0,0 +1,22 @@
+//===-- AMDGPURegisterInfo.td - AMDGPU register info -------*- tablegen -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// Tablegen register definitions common to all hw codegen targets.
+//
+//===----------------------------------------------------------------------===//
+
+let Namespace = "AMDGPU" in {
+  def sel_x : SubRegIndex;
+  def sel_y : SubRegIndex;
+  def sel_z : SubRegIndex;
+  def sel_w : SubRegIndex;
+}
+
+include "R600RegisterInfo.td"
+include "SIRegisterInfo.td"
diff --git a/lib/Target/R600/AMDGPUSubtarget.cpp b/lib/Target/R600/AMDGPUSubtarget.cpp
new file mode 100644
index 0000000000..0f356a1c3f
--- /dev/null
+++ b/lib/Target/R600/AMDGPUSubtarget.cpp
@@ -0,0 +1,87 @@
+//===-- AMDGPUSubtarget.cpp - AMDGPU Subtarget Information ----------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Implements the AMDGPU specific subclass of TargetSubtarget.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUSubtarget.h"
+
+using namespace llvm;
+
+#define GET_SUBTARGETINFO_ENUM
+#define GET_SUBTARGETINFO_TARGET_DESC
+#define GET_SUBTARGETINFO_CTOR
+#include "AMDGPUGenSubtargetInfo.inc"
+
+AMDGPUSubtarget::AMDGPUSubtarget(StringRef TT, StringRef CPU, StringRef FS) :
+  AMDGPUGenSubtargetInfo(TT, CPU, FS), DumpCode(false) {
+    InstrItins = getInstrItineraryForCPU(CPU);
+
+  memset(CapsOverride, 0, sizeof(*CapsOverride)
+      * AMDGPUDeviceInfo::MaxNumberCapabilities);
+  // Default card
+  StringRef GPU = CPU;
+  Is64bit = false;
+  DefaultSize[0] = 64;
+  DefaultSize[1] = 1;
+  DefaultSize[2] = 1;
+  ParseSubtargetFeatures(GPU, FS);
+  DevName = GPU;
+  Device = AMDGPUDeviceInfo::getDeviceFromName(DevName, this, Is64bit);
+}
+
+AMDGPUSubtarget::~AMDGPUSubtarget() {
+  delete Device;
+}
+
+bool
+AMDGPUSubtarget::isOverride(AMDGPUDeviceInfo::Caps caps) const {
+  assert(caps < AMDGPUDeviceInfo::MaxNumberCapabilities &&
+      "Caps index is out of bounds!");
+  return CapsOverride[caps];
+}
+bool
+AMDGPUSubtarget::is64bit() const  {
+  return Is64bit;
+}
+bool
+AMDGPUSubtarget::isTargetELF() const {
+  return false;
+}
+size_t
+AMDGPUSubtarget::getDefaultSize(uint32_t dim) const {
+  if (dim > 3) {
+    return 1;
+  } else {
+    return DefaultSize[dim];
+  }
+}
+
+std::string
+AMDGPUSubtarget::getDataLayout() const {
+    if (!Device) {
+        return std::string("e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16"
+                "-i32:32:32-i64:64:64-f32:32:32-f64:64:64-f80:32:32"
+                "-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64"
+                "-v96:128:128-v128:128:128-v192:256:256-v256:256:256"
+                "-v512:512:512-v1024:1024:1024-v2048:2048:2048-a0:0:64");
+    }
+    return Device->getDataLayout();
+}
+
+std::string
+AMDGPUSubtarget::getDeviceName() const {
+  return DevName;
+}
+const AMDGPUDevice *
+AMDGPUSubtarget::device() const {
+  return Device;
+}
diff --git a/lib/Target/R600/AMDGPUSubtarget.h b/lib/Target/R600/AMDGPUSubtarget.h
new file mode 100644
index 0000000000..cab7884ea3
--- /dev/null
+++ b/lib/Target/R600/AMDGPUSubtarget.h
@@ -0,0 +1,65 @@
+//=====-- AMDGPUSubtarget.h - Define Subtarget for the AMDIL ---*- C++ -*-====//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief AMDGPU specific subclass of TargetSubtarget.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef AMDGPUSUBTARGET_H
+#define AMDGPUSUBTARGET_H
+#include "AMDILDevice.h"
+#include "llvm/ADT/StringExtras.h"
+#include "llvm/ADT/StringRef.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
+
+#define GET_SUBTARGETINFO_HEADER
+#include "AMDGPUGenSubtargetInfo.inc"
+
+#define MAX_CB_SIZE (1 << 16)
+
+namespace llvm {
+
+class AMDGPUSubtarget : public AMDGPUGenSubtargetInfo {
+private:
+  bool CapsOverride[AMDGPUDeviceInfo::MaxNumberCapabilities];
+  const AMDGPUDevice *Device;
+  size_t DefaultSize[3];
+  std::string DevName;
+  bool Is64bit;
+  bool Is32on64bit;
+  bool DumpCode;
+  bool R600ALUInst;
+
+  InstrItineraryData InstrItins;
+
+public:
+  AMDGPUSubtarget(StringRef TT, StringRef CPU, StringRef FS);
+  virtual ~AMDGPUSubtarget();
+
+  const InstrItineraryData &getInstrItineraryData() const { return InstrItins; }
+  virtual void ParseSubtargetFeatures(llvm::StringRef CPU, llvm::StringRef FS);
+
+  bool isOverride(AMDGPUDeviceInfo::Caps) const;
+  bool is64bit() const;
+
+  // Helper functions to simplify if statements
+  bool isTargetELF() const;
+  const AMDGPUDevice* device() const;
+  std::string getDataLayout() const;
+  std::string getDeviceName() const;
+  virtual size_t getDefaultSize(uint32_t dim) const;
+  bool dumpCode() const { return DumpCode; }
+  bool r600ALUEncoding() const { return R600ALUInst; }
+
+};
+
+} // End namespace llvm
+
+#endif // AMDGPUSUBTARGET_H
diff --git a/lib/Target/R600/AMDGPUTargetMachine.cpp b/lib/Target/R600/AMDGPUTargetMachine.cpp
new file mode 100644
index 0000000000..98a3064f7e
--- /dev/null
+++ b/lib/Target/R600/AMDGPUTargetMachine.cpp
@@ -0,0 +1,141 @@
+//===-- AMDGPUTargetMachine.cpp - TargetMachine for hw codegen targets-----===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief The AMDGPU target machine contains all of the hardware specific
+/// information  needed to emit code for R600 and SI GPUs.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUTargetMachine.h"
+#include "AMDGPU.h"
+#include "R600ISelLowering.h"
+#include "R600InstrInfo.h"
+#include "SIISelLowering.h"
+#include "SIInstrInfo.h"
+#include "llvm/Analysis/Passes.h"
+#include "llvm/Analysis/Verifier.h"
+#include "llvm/CodeGen/MachineFunctionAnalysis.h"
+#include "llvm/CodeGen/MachineModuleInfo.h"
+#include "llvm/CodeGen/Passes.h"
+#include "llvm/MC/MCAsmInfo.h"
+#include "llvm/PassManager.h"
+#include "llvm/Support/TargetRegistry.h"
+#include "llvm/Support/raw_os_ostream.h"
+#include "llvm/Transforms/IPO.h"
+#include "llvm/Transforms/Scalar.h"
+#include <llvm/CodeGen/Passes.h>
+
+using namespace llvm;
+
+extern "C" void LLVMInitializeR600Target() {
+  // Register the target
+  RegisterTargetMachine<AMDGPUTargetMachine> X(TheAMDGPUTarget);
+}
+
+AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, StringRef TT,
+    StringRef CPU, StringRef FS,
+  TargetOptions Options,
+  Reloc::Model RM, CodeModel::Model CM,
+  CodeGenOpt::Level OptLevel
+)
+:
+  LLVMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OptLevel),
+  Subtarget(TT, CPU, FS),
+  Layout(Subtarget.getDataLayout()),
+  FrameLowering(TargetFrameLowering::StackGrowsUp,
+      Subtarget.device()->getStackAlignment(), 0),
+  IntrinsicInfo(this),
+  InstrItins(&Subtarget.getInstrItineraryData()) {
+  // TLInfo uses InstrInfo so it must be initialized after.
+  if (Subtarget.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX) {
+    InstrInfo = new R600InstrInfo(*this);
+    TLInfo = new R600TargetLowering(*this);
+  } else {
+    InstrInfo = new SIInstrInfo(*this);
+    TLInfo = new SITargetLowering(*this);
+  }
+}
+
+AMDGPUTargetMachine::~AMDGPUTargetMachine() {
+}
+
+namespace {
+class AMDGPUPassConfig : public TargetPassConfig {
+public:
+  AMDGPUPassConfig(AMDGPUTargetMachine *TM, PassManagerBase &PM)
+    : TargetPassConfig(TM, PM) {}
+
+  AMDGPUTargetMachine &getAMDGPUTargetMachine() const {
+    return getTM<AMDGPUTargetMachine>();
+  }
+
+  virtual bool addPreISel();
+  virtual bool addInstSelector();
+  virtual bool addPreRegAlloc();
+  virtual bool addPostRegAlloc();
+  virtual bool addPreSched2();
+  virtual bool addPreEmitPass();
+};
+} // End of anonymous namespace
+
+TargetPassConfig *AMDGPUTargetMachine::createPassConfig(PassManagerBase &PM) {
+  return new AMDGPUPassConfig(this, PM);
+}
+
+bool
+AMDGPUPassConfig::addPreISel() {
+  return false;
+}
+
+bool AMDGPUPassConfig::addInstSelector() {
+  addPass(createAMDGPUPeepholeOpt(*TM));
+  addPass(createAMDGPUISelDag(getAMDGPUTargetMachine()));
+  return false;
+}
+
+bool AMDGPUPassConfig::addPreRegAlloc() {
+  const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
+
+  if (ST.device()->getGeneration() > AMDGPUDeviceInfo::HD6XXX) {
+    addPass(createSIAssignInterpRegsPass(*TM));
+  }
+  addPass(createAMDGPUConvertToISAPass(*TM));
+  if (ST.device()->getGeneration() > AMDGPUDeviceInfo::HD6XXX) {
+    addPass(createSIFixSGPRLivenessPass(*TM));
+  }
+  return false;
+}
+
+bool AMDGPUPassConfig::addPostRegAlloc() {
+  return false;
+}
+
+bool AMDGPUPassConfig::addPreSched2() {
+
+  addPass(&IfConverterID);
+  return false;
+}
+
+bool AMDGPUPassConfig::addPreEmitPass() {
+  addPass(createAMDGPUCFGPreparationPass(*TM));
+  addPass(createAMDGPUCFGStructurizerPass(*TM));
+
+  const AMDGPUSubtarget &ST = TM->getSubtarget<AMDGPUSubtarget>();
+  if (ST.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX) {
+    addPass(createR600ExpandSpecialInstrsPass(*TM));
+    addPass(&FinalizeMachineBundlesID);
+  } else {
+    addPass(createSILowerLiteralConstantsPass(*TM));
+    addPass(createSILowerControlFlowPass(*TM));
+  }
+
+  return false;
+}
+
diff --git a/lib/Target/R600/AMDGPUTargetMachine.h b/lib/Target/R600/AMDGPUTargetMachine.h
new file mode 100644
index 0000000000..399e55c8ad
--- /dev/null
+++ b/lib/Target/R600/AMDGPUTargetMachine.h
@@ -0,0 +1,70 @@
+//===-- AMDGPUTargetMachine.h - AMDGPU TargetMachine Interface --*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief The AMDGPU TargetMachine interface definition for hw codgen targets.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef AMDGPU_TARGET_MACHINE_H
+#define AMDGPU_TARGET_MACHINE_H
+
+#include "AMDGPUInstrInfo.h"
+#include "AMDGPUSubtarget.h"
+#include "AMDILFrameLowering.h"
+#include "AMDILIntrinsicInfo.h"
+#include "R600ISelLowering.h"
+#include "llvm/ADT/OwningPtr.h"
+#include "llvm/DataLayout.h"
+
+namespace llvm {
+
+MCAsmInfo* createMCAsmInfo(const Target &T, StringRef TT);
+
+class AMDGPUTargetMachine : public LLVMTargetMachine {
+
+  AMDGPUSubtarget Subtarget;
+  const DataLayout Layout;
+  AMDGPUFrameLowering FrameLowering;
+  AMDGPUIntrinsicInfo IntrinsicInfo;
+  const AMDGPUInstrInfo * InstrInfo;
+  AMDGPUTargetLowering * TLInfo;
+  const InstrItineraryData* InstrItins;
+
+public:
+   AMDGPUTargetMachine(const Target &T, StringRef TT, StringRef FS,
+                       StringRef CPU,
+                       TargetOptions Options,
+                       Reloc::Model RM, CodeModel::Model CM,
+                       CodeGenOpt::Level OL);
+   ~AMDGPUTargetMachine();
+   virtual const AMDGPUFrameLowering* getFrameLowering() const {
+     return &FrameLowering;
+   }
+   virtual const AMDGPUIntrinsicInfo* getIntrinsicInfo() const {
+     return &IntrinsicInfo;
+   }
+   virtual const AMDGPUInstrInfo *getInstrInfo() const {return InstrInfo;}
+   virtual const AMDGPUSubtarget *getSubtargetImpl() const {return &Subtarget; }
+   virtual const AMDGPURegisterInfo *getRegisterInfo() const {
+      return &InstrInfo->getRegisterInfo();
+   }
+   virtual AMDGPUTargetLowering * getTargetLowering() const {
+      return TLInfo;
+   }
+   virtual const InstrItineraryData* getInstrItineraryData() const {
+      return InstrItins;
+   }
+   virtual const DataLayout* getDataLayout() const { return &Layout; }
+   virtual TargetPassConfig *createPassConfig(PassManagerBase &PM);
+};
+
+} // End namespace llvm
+
+#endif // AMDGPU_TARGET_MACHINE_H
diff --git a/lib/Target/R600/AMDIL.h b/lib/Target/R600/AMDIL.h
new file mode 100644
index 0000000000..4e577dc234
--- /dev/null
+++ b/lib/Target/R600/AMDIL.h
@@ -0,0 +1,106 @@
+//===-- AMDIL.h - Top-level interface for AMDIL representation --*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+/// This file contains the entry points for global functions defined in the LLVM
+/// AMDGPU back-end.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef AMDIL_H
+#define AMDIL_H
+
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/Target/TargetMachine.h"
+
+#define ARENA_SEGMENT_RESERVED_UAVS 12
+#define DEFAULT_ARENA_UAV_ID 8
+#define DEFAULT_RAW_UAV_ID 7
+#define GLOBAL_RETURN_RAW_UAV_ID 11
+#define HW_MAX_NUM_CB 8
+#define MAX_NUM_UNIQUE_UAVS 8
+#define OPENCL_MAX_NUM_ATOMIC_COUNTERS 8
+#define OPENCL_MAX_READ_IMAGES 128
+#define OPENCL_MAX_WRITE_IMAGES 8
+#define OPENCL_MAX_SAMPLERS 16
+
+// The next two values can never be zero, as zero is the ID that is
+// used to assert against.
+#define DEFAULT_LDS_ID     1
+#define DEFAULT_GDS_ID     1
+#define DEFAULT_SCRATCH_ID 1
+#define DEFAULT_VEC_SLOTS  8
+
+#define OCL_DEVICE_RV710        0x0001
+#define OCL_DEVICE_RV730        0x0002
+#define OCL_DEVICE_RV770        0x0004
+#define OCL_DEVICE_CEDAR        0x0008
+#define OCL_DEVICE_REDWOOD      0x0010
+#define OCL_DEVICE_JUNIPER      0x0020
+#define OCL_DEVICE_CYPRESS      0x0040
+#define OCL_DEVICE_CAICOS       0x0080
+#define OCL_DEVICE_TURKS        0x0100
+#define OCL_DEVICE_BARTS        0x0200
+#define OCL_DEVICE_CAYMAN       0x0400
+#define OCL_DEVICE_ALL          0x3FFF
+
+/// The number of function ID's that are reserved for 
+/// internal compiler usage.
+const unsigned int RESERVED_FUNCS = 1024;
+
+namespace llvm {
+class AMDGPUInstrPrinter;
+class FunctionPass;
+class MCAsmInfo;
+class raw_ostream;
+class Target;
+class TargetMachine;
+
+// Instruction selection passes.
+FunctionPass*
+  createAMDGPUISelDag(TargetMachine &TM);
+FunctionPass*
+  createAMDGPUPeepholeOpt(TargetMachine &TM);
+
+// Pre emit passes.
+FunctionPass*
+  createAMDGPUCFGPreparationPass(TargetMachine &TM);
+FunctionPass*
+  createAMDGPUCFGStructurizerPass(TargetMachine &TM);
+
+extern Target TheAMDGPUTarget;
+} // end namespace llvm;
+
+// Include device information enumerations
+#include "AMDILDeviceInfo.h"
+
+namespace llvm {
+/// OpenCL uses address spaces to differentiate between
+/// various memory regions on the hardware. On the CPU
+/// all of the address spaces point to the same memory,
+/// however on the GPU, each address space points to
+/// a seperate piece of memory that is unique from other
+/// memory locations.
+namespace AMDGPUAS {
+enum AddressSpaces {
+  PRIVATE_ADDRESS  = 0, ///< Address space for private memory.
+  GLOBAL_ADDRESS   = 1, ///< Address space for global memory (RAT0, VTX0).
+  CONSTANT_ADDRESS = 2, ///< Address space for constant memory.
+  LOCAL_ADDRESS    = 3, ///< Address space for local memory.
+  REGION_ADDRESS   = 4, ///< Address space for region memory.
+  ADDRESS_NONE     = 5, ///< Address space for unknown memory.
+  PARAM_D_ADDRESS  = 6, ///< Address space for direct addressible parameter memory (CONST0)
+  PARAM_I_ADDRESS  = 7, ///< Address space for indirect addressible parameter memory (VTX1)
+  USER_SGPR_ADDRESS = 8, ///< Address space for USER_SGPRS on SI
+  LAST_ADDRESS     = 9
+};
+
+} // namespace AMDGPUAS
+
+} // end namespace llvm
+#endif // AMDIL_H
diff --git a/lib/Target/R600/AMDIL7XXDevice.cpp b/lib/Target/R600/AMDIL7XXDevice.cpp
new file mode 100644
index 0000000000..ea6ac34f57
--- /dev/null
+++ b/lib/Target/R600/AMDIL7XXDevice.cpp
@@ -0,0 +1,115 @@
+//===-- AMDIL7XXDevice.cpp - Device Info for 7XX GPUs ---------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+// \file
+//==-----------------------------------------------------------------------===//
+#include "AMDIL7XXDevice.h"
+#include "AMDGPUSubtarget.h"
+#include "AMDILDevice.h"
+
+using namespace llvm;
+
+AMDGPU7XXDevice::AMDGPU7XXDevice(AMDGPUSubtarget *ST) : AMDGPUDevice(ST) {
+  setCaps();
+  std::string name = mSTM->getDeviceName();
+  if (name == "rv710") {
+    DeviceFlag = OCL_DEVICE_RV710;
+  } else if (name == "rv730") {
+    DeviceFlag = OCL_DEVICE_RV730;
+  } else {
+    DeviceFlag = OCL_DEVICE_RV770;
+  }
+}
+
+AMDGPU7XXDevice::~AMDGPU7XXDevice() {
+}
+
+void AMDGPU7XXDevice::setCaps() {
+  mSWBits.set(AMDGPUDeviceInfo::LocalMem);
+}
+
+size_t AMDGPU7XXDevice::getMaxLDSSize() const {
+  if (usesHardware(AMDGPUDeviceInfo::LocalMem)) {
+    return MAX_LDS_SIZE_700;
+  }
+  return 0;
+}
+
+size_t AMDGPU7XXDevice::getWavefrontSize() const {
+  return AMDGPUDevice::HalfWavefrontSize;
+}
+
+uint32_t AMDGPU7XXDevice::getGeneration() const {
+  return AMDGPUDeviceInfo::HD4XXX;
+}
+
+uint32_t AMDGPU7XXDevice::getResourceID(uint32_t DeviceID) const {
+  switch (DeviceID) {
+  default:
+    assert(0 && "ID type passed in is unknown!");
+    break;
+  case GLOBAL_ID:
+  case CONSTANT_ID:
+  case RAW_UAV_ID:
+  case ARENA_UAV_ID:
+    break;
+  case LDS_ID:
+    if (usesHardware(AMDGPUDeviceInfo::LocalMem)) {
+      return DEFAULT_LDS_ID;
+    }
+    break;
+  case SCRATCH_ID:
+    if (usesHardware(AMDGPUDeviceInfo::PrivateMem)) {
+      return DEFAULT_SCRATCH_ID;
+    }
+    break;
+  case GDS_ID:
+    assert(0 && "GDS UAV ID is not supported on this chip");
+    if (usesHardware(AMDGPUDeviceInfo::RegionMem)) {
+      return DEFAULT_GDS_ID;
+    }
+    break;
+  };
+
+  return 0;
+}
+
+uint32_t AMDGPU7XXDevice::getMaxNumUAVs() const {
+  return 1;
+}
+
+AMDGPU770Device::AMDGPU770Device(AMDGPUSubtarget *ST): AMDGPU7XXDevice(ST) {
+  setCaps();
+}
+
+AMDGPU770Device::~AMDGPU770Device() {
+}
+
+void AMDGPU770Device::setCaps() {
+  if (mSTM->isOverride(AMDGPUDeviceInfo::DoubleOps)) {
+    mSWBits.set(AMDGPUDeviceInfo::FMA);
+    mHWBits.set(AMDGPUDeviceInfo::DoubleOps);
+  }
+  mSWBits.set(AMDGPUDeviceInfo::BarrierDetect);
+  mHWBits.reset(AMDGPUDeviceInfo::LongOps);
+  mSWBits.set(AMDGPUDeviceInfo::LongOps);
+  mSWBits.set(AMDGPUDeviceInfo::LocalMem);
+}
+
+size_t AMDGPU770Device::getWavefrontSize() const {
+  return AMDGPUDevice::WavefrontSize;
+}
+
+AMDGPU710Device::AMDGPU710Device(AMDGPUSubtarget *ST) : AMDGPU7XXDevice(ST) {
+}
+
+AMDGPU710Device::~AMDGPU710Device() {
+}
+
+size_t AMDGPU710Device::getWavefrontSize() const {
+  return AMDGPUDevice::QuarterWavefrontSize;
+}
diff --git a/lib/Target/R600/AMDIL7XXDevice.h b/lib/Target/R600/AMDIL7XXDevice.h
new file mode 100644
index 0000000000..1cf4ca415a
--- /dev/null
+++ b/lib/Target/R600/AMDIL7XXDevice.h
@@ -0,0 +1,72 @@
+//==-- AMDIL7XXDevice.h - Define 7XX Device Device for AMDIL ---*- C++ -*--===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+/// \file
+/// \brief Interface for the subtarget data classes.
+///
+/// This file will define the interface that each generation needs to
+/// implement in order to correctly answer queries on the capabilities of the
+/// specific hardware.
+//===----------------------------------------------------------------------===//
+#ifndef AMDIL7XXDEVICEIMPL_H
+#define AMDIL7XXDEVICEIMPL_H
+#include "AMDILDevice.h"
+
+namespace llvm {
+class AMDGPUSubtarget;
+
+//===----------------------------------------------------------------------===//
+// 7XX generation of devices and their respective sub classes
+//===----------------------------------------------------------------------===//
+
+/// \brief The AMDGPU7XXDevice class represents the generic 7XX device.
+///
+/// All 7XX devices are derived from this class. The AMDGPU7XX device will only
+/// support the minimal features that are required to be considered OpenCL 1.0
+/// compliant and nothing more.
+class AMDGPU7XXDevice : public AMDGPUDevice {
+public:
+  AMDGPU7XXDevice(AMDGPUSubtarget *ST);
+  virtual ~AMDGPU7XXDevice();
+  virtual size_t getMaxLDSSize() const;
+  virtual size_t getWavefrontSize() const;
+  virtual uint32_t getGeneration() const;
+  virtual uint32_t getResourceID(uint32_t DeviceID) const;
+  virtual uint32_t getMaxNumUAVs() const;
+
+protected:
+  virtual void setCaps();
+};
+
+/// \brief The AMDGPU770Device class represents the RV770 chip and it's
+/// derivative cards.
+///
+/// The difference between this device and the base class is this device device
+/// adds support for double precision and has a larger wavefront size.
+class AMDGPU770Device : public AMDGPU7XXDevice {
+public:
+  AMDGPU770Device(AMDGPUSubtarget *ST);
+  virtual ~AMDGPU770Device();
+  virtual size_t getWavefrontSize() const;
+private:
+  virtual void setCaps();
+};
+
+/// \brief The AMDGPU710Device class derives from the 7XX base class.
+///
+/// This class is a smaller derivative, so we need to overload some of the
+/// functions in order to correctly specify this information.
+class AMDGPU710Device : public AMDGPU7XXDevice {
+public:
+  AMDGPU710Device(AMDGPUSubtarget *ST);
+  virtual ~AMDGPU710Device();
+  virtual size_t getWavefrontSize() const;
+};
+
+} // namespace llvm
+#endif // AMDILDEVICEIMPL_H
diff --git a/lib/Target/R600/AMDILBase.td b/lib/Target/R600/AMDILBase.td
new file mode 100644
index 0000000000..c12cedcf7f
--- /dev/null
+++ b/lib/Target/R600/AMDILBase.td
@@ -0,0 +1,85 @@
+//===- AMDIL.td - AMDIL Target Machine -------------*- tablegen -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+// Target-independent interfaces which we are implementing
+//===----------------------------------------------------------------------===//
+
+include "llvm/Target/Target.td"
+
+// Dummy Instruction itineraries for pseudo instructions
+def ALU_NULL : FuncUnit;
+def NullALU : InstrItinClass;
+
+//===----------------------------------------------------------------------===//
+// AMDIL Subtarget features.
+//===----------------------------------------------------------------------===//
+def FeatureFP64     : SubtargetFeature<"fp64",
+        "CapsOverride[AMDGPUDeviceInfo::DoubleOps]",
+        "true",
+        "Enable 64bit double precision operations">;
+def FeatureByteAddress    : SubtargetFeature<"byte_addressable_store",
+        "CapsOverride[AMDGPUDeviceInfo::ByteStores]",
+        "true",
+        "Enable byte addressable stores">;
+def FeatureBarrierDetect : SubtargetFeature<"barrier_detect",
+        "CapsOverride[AMDGPUDeviceInfo::BarrierDetect]",
+        "true",
+        "Enable duplicate barrier detection(HD5XXX or later).">;
+def FeatureImages : SubtargetFeature<"images",
+        "CapsOverride[AMDGPUDeviceInfo::Images]",
+        "true",
+        "Enable image functions">;
+def FeatureMultiUAV : SubtargetFeature<"multi_uav",
+        "CapsOverride[AMDGPUDeviceInfo::MultiUAV]",
+        "true",
+        "Generate multiple UAV code(HD5XXX family or later)">;
+def FeatureMacroDB : SubtargetFeature<"macrodb",
+        "CapsOverride[AMDGPUDeviceInfo::MacroDB]",
+        "true",
+        "Use internal macrodb, instead of macrodb in driver">;
+def FeatureNoAlias : SubtargetFeature<"noalias",
+        "CapsOverride[AMDGPUDeviceInfo::NoAlias]",
+        "true",
+        "assert that all kernel argument pointers are not aliased">;
+def FeatureNoInline : SubtargetFeature<"no-inline",
+        "CapsOverride[AMDGPUDeviceInfo::NoInline]",
+        "true",
+        "specify whether to not inline functions">;
+
+def Feature64BitPtr : SubtargetFeature<"64BitPtr",
+        "Is64bit",
+        "false",
+        "Specify if 64bit addressing should be used.">;
+
+def Feature32on64BitPtr : SubtargetFeature<"64on32BitPtr",
+        "Is32on64bit",
+        "false",
+        "Specify if 64bit sized pointers with 32bit addressing should be used.">;
+def FeatureDebug : SubtargetFeature<"debug",
+        "CapsOverride[AMDGPUDeviceInfo::Debug]",
+        "true",
+        "Debug mode is enabled, so disable hardware accelerated address spaces.">;
+def FeatureDumpCode : SubtargetFeature <"DumpCode",
+        "DumpCode",
+        "true",
+        "Dump MachineInstrs in the CodeEmitter">;
+
+def FeatureR600ALUInst : SubtargetFeature<"R600ALUInst",
+        "R600ALUInst",
+        "false",
+        "Older version of ALU instructions encoding.">;
+
+
+//===----------------------------------------------------------------------===//
+// Register File, Calling Conv, Instruction Descriptions
+//===----------------------------------------------------------------------===//
+
+
+include "AMDILRegisterInfo.td"
+include "AMDILInstrInfo.td"
+
diff --git a/lib/Target/R600/AMDILCFGStructurizer.cpp b/lib/Target/R600/AMDILCFGStructurizer.cpp
new file mode 100644
index 0000000000..1f276dc570
--- /dev/null
+++ b/lib/Target/R600/AMDILCFGStructurizer.cpp
@@ -0,0 +1,3049 @@
+//===-- AMDILCFGStructurizer.cpp - CFG Structurizer -----------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//==-----------------------------------------------------------------------===//
+
+#define DEBUGME 0
+#define DEBUG_TYPE "structcfg"
+
+#include "AMDGPUInstrInfo.h"
+#include "AMDIL.h"
+#include "llvm/ADT/SCCIterator.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/DominatorInternals.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/CodeGen/MachinePostDominators.h"
+#include "llvm/CodeGen/MachineDominators.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFunctionAnalysis.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineJumpTableInfo.h"
+#include "llvm/CodeGen/MachineLoopInfo.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/Target/TargetInstrInfo.h"
+
+using namespace llvm;
+
+// TODO: move-begin.
+
+//===----------------------------------------------------------------------===//
+//
+// Statistics for CFGStructurizer.
+//
+//===----------------------------------------------------------------------===//
+
+STATISTIC(numSerialPatternMatch,    "CFGStructurizer number of serial pattern "
+    "matched");
+STATISTIC(numIfPatternMatch,        "CFGStructurizer number of if pattern "
+    "matched");
+STATISTIC(numLoopbreakPatternMatch, "CFGStructurizer number of loop-break "
+    "pattern matched");
+STATISTIC(numLoopcontPatternMatch,  "CFGStructurizer number of loop-continue "
+    "pattern matched");
+STATISTIC(numLoopPatternMatch,      "CFGStructurizer number of loop pattern "
+    "matched");
+STATISTIC(numClonedBlock,           "CFGStructurizer cloned blocks");
+STATISTIC(numClonedInstr,           "CFGStructurizer cloned instructions");
+
+//===----------------------------------------------------------------------===//
+//
+// Miscellaneous utility for CFGStructurizer.
+//
+//===----------------------------------------------------------------------===//
+namespace llvmCFGStruct {
+#define SHOWNEWINSTR(i) \
+  if (DEBUGME) errs() << "New instr: " << *i << "\n"
+
+#define SHOWNEWBLK(b, msg) \
+if (DEBUGME) { \
+  errs() << msg << "BB" << b->getNumber() << "size " << b->size(); \
+  errs() << "\n"; \
+}
+
+#define SHOWBLK_DETAIL(b, msg) \
+if (DEBUGME) { \
+  if (b) { \
+  errs() << msg << "BB" << b->getNumber() << "size " << b->size(); \
+  b->print(errs()); \
+  errs() << "\n"; \
+  } \
+}
+
+#define INVALIDSCCNUM -1
+#define INVALIDREGNUM 0
+
+template<class LoopinfoT>
+void PrintLoopinfo(const LoopinfoT &LoopInfo, llvm::raw_ostream &OS) {
+  for (typename LoopinfoT::iterator iter = LoopInfo.begin(),
+       iterEnd = LoopInfo.end();
+       iter != iterEnd; ++iter) {
+    (*iter)->print(OS, 0);
+  }
+}
+
+template<class NodeT>
+void ReverseVector(SmallVector<NodeT *, DEFAULT_VEC_SLOTS> &Src) {
+  size_t sz = Src.size();
+  for (size_t i = 0; i < sz/2; ++i) {
+    NodeT *t = Src[i];
+    Src[i] = Src[sz - i - 1];
+    Src[sz - i - 1] = t;
+  }
+}
+
+} //end namespace llvmCFGStruct
+
+//===----------------------------------------------------------------------===//
+//
+// supporting data structure for CFGStructurizer
+//
+//===----------------------------------------------------------------------===//
+
+namespace llvmCFGStruct {
+template<class PassT>
+struct CFGStructTraits {
+};
+
+template <class InstrT>
+class BlockInformation {
+public:
+  bool isRetired;
+  int  sccNum;
+  //SmallVector<InstrT*, DEFAULT_VEC_SLOTS> succInstr;
+  //Instructions defining the corresponding successor.
+  BlockInformation() : isRetired(false), sccNum(INVALIDSCCNUM) {}
+};
+
+template <class BlockT, class InstrT, class RegiT>
+class LandInformation {
+public:
+  BlockT *landBlk;
+  std::set<RegiT> breakInitRegs;  //Registers that need to "reg = 0", before
+                                  //WHILELOOP(thisloop) init before entering
+                                  //thisloop.
+  std::set<RegiT> contInitRegs;   //Registers that need to "reg = 0", after
+                                  //WHILELOOP(thisloop) init after entering
+                                  //thisloop.
+  std::set<RegiT> endbranchInitRegs; //Init before entering this loop, at loop
+                                     //land block, branch cond on this reg.
+  std::set<RegiT> breakOnRegs;       //registers that need to "if (reg) break
+                                     //endif" after ENDLOOP(thisloop) break
+                                     //outerLoopOf(thisLoop).
+  std::set<RegiT> contOnRegs;       //registers that need to "if (reg) continue
+                                    //endif" after ENDLOOP(thisloop) continue on
+                                    //outerLoopOf(thisLoop).
+  LandInformation() : landBlk(NULL) {}
+};
+
+} //end of namespace llvmCFGStruct
+
+//===----------------------------------------------------------------------===//
+//
+// CFGStructurizer
+//
+//===----------------------------------------------------------------------===//
+
+namespace llvmCFGStruct {
+// bixia TODO: port it to BasicBlock, not just MachineBasicBlock.
+template<class PassT>
+class  CFGStructurizer {
+public:
+  typedef enum {
+    Not_SinglePath = 0,
+    SinglePath_InPath = 1,
+    SinglePath_NotInPath = 2
+  } PathToKind;
+
+public:
+  typedef typename PassT::InstructionType         InstrT;
+  typedef typename PassT::FunctionType            FuncT;
+  typedef typename PassT::DominatortreeType       DomTreeT;
+  typedef typename PassT::PostDominatortreeType   PostDomTreeT;
+  typedef typename PassT::DomTreeNodeType         DomTreeNodeT;
+  typedef typename PassT::LoopinfoType            LoopInfoT;
+
+  typedef GraphTraits<FuncT *>                    FuncGTraits;
+  //typedef FuncGTraits::nodes_iterator BlockIterator;
+  typedef typename FuncT::iterator                BlockIterator;
+
+  typedef typename FuncGTraits::NodeType          BlockT;
+  typedef GraphTraits<BlockT *>                   BlockGTraits;
+  typedef GraphTraits<Inverse<BlockT *> >         InvBlockGTraits;
+  //typedef BlockGTraits::succ_iterator InstructionIterator;
+  typedef typename BlockT::iterator               InstrIterator;
+
+  typedef CFGStructTraits<PassT>                  CFGTraits;
+  typedef BlockInformation<InstrT>                BlockInfo;
+  typedef std::map<BlockT *, BlockInfo *>         BlockInfoMap;
+
+  typedef int                                     RegiT;
+  typedef typename PassT::LoopType                LoopT;
+  typedef LandInformation<BlockT, InstrT, RegiT>  LoopLandInfo;
+        typedef std::map<LoopT *, LoopLandInfo *> LoopLandInfoMap;
+        //landing info for loop break
+  typedef SmallVector<BlockT *, 32>               BlockTSmallerVector;
+
+public:
+  CFGStructurizer();
+  ~CFGStructurizer();
+
+  /// Perform the CFG structurization
+  bool run(FuncT &Func, PassT &Pass, const AMDGPURegisterInfo *tri);
+
+  /// Perform the CFG preparation
+  bool prepare(FuncT &Func, PassT &Pass, const AMDGPURegisterInfo *tri);
+
+private:
+  void reversePredicateSetter(typename BlockT::iterator);
+  void   orderBlocks();
+  void   printOrderedBlocks(llvm::raw_ostream &OS);
+  int patternMatch(BlockT *CurBlock);
+  int patternMatchGroup(BlockT *CurBlock);
+
+  int serialPatternMatch(BlockT *CurBlock);
+  int ifPatternMatch(BlockT *CurBlock);
+  int switchPatternMatch(BlockT *CurBlock);
+  int loopendPatternMatch(BlockT *CurBlock);
+  int loopPatternMatch(BlockT *CurBlock);
+
+  int loopbreakPatternMatch(LoopT *LoopRep, BlockT *LoopHeader);
+  int loopcontPatternMatch(LoopT *LoopRep, BlockT *LoopHeader);
+  //int loopWithoutBreak(BlockT *);
+
+  void handleLoopbreak (BlockT *ExitingBlock, LoopT *ExitingLoop,
+                        BlockT *ExitBlock, LoopT *exitLoop, BlockT *landBlock);
+  void handleLoopcontBlock(BlockT *ContingBlock, LoopT *contingLoop,
+                           BlockT *ContBlock, LoopT *contLoop);
+  bool isSameloopDetachedContbreak(BlockT *Src1Block, BlockT *Src2Block);
+  int handleJumpintoIf(BlockT *HeadBlock, BlockT *TrueBlock,
+                       BlockT *FalseBlock);
+  int handleJumpintoIfImp(BlockT *HeadBlock, BlockT *TrueBlock,
+                          BlockT *FalseBlock);
+  int improveSimpleJumpintoIf(BlockT *HeadBlock, BlockT *TrueBlock,
+                              BlockT *FalseBlock, BlockT **LandBlockPtr);
+  void showImproveSimpleJumpintoIf(BlockT *HeadBlock, BlockT *TrueBlock,
+                                   BlockT *FalseBlock, BlockT *LandBlock,
+                                   bool Detail = false);
+  PathToKind singlePathTo(BlockT *SrcBlock, BlockT *DstBlock,
+                          bool AllowSideEntry = true);
+  BlockT *singlePathEnd(BlockT *srcBlock, BlockT *DstBlock,
+                        bool AllowSideEntry = true);
+  int cloneOnSideEntryTo(BlockT *PreBlock, BlockT *SrcBlock, BlockT *DstBlock);
+  void mergeSerialBlock(BlockT *DstBlock, BlockT *srcBlock);
+
+  void mergeIfthenelseBlock(InstrT *BranchInstr, BlockT *CurBlock,
+                            BlockT *TrueBlock, BlockT *FalseBlock,
+                            BlockT *LandBlock);
+  void mergeLooplandBlock(BlockT *DstBlock, LoopLandInfo *LoopLand);
+  void mergeLoopbreakBlock(BlockT *ExitingBlock, BlockT *ExitBlock,
+                           BlockT *ExitLandBlock, RegiT SetReg);
+  void settleLoopcontBlock(BlockT *ContingBlock, BlockT *ContBlock,
+                           RegiT SetReg);
+  BlockT *relocateLoopcontBlock(LoopT *ParentLoopRep, LoopT *LoopRep,
+                                std::set<BlockT*> &ExitBlockSet,
+                                BlockT *ExitLandBlk);
+  BlockT *addLoopEndbranchBlock(LoopT *LoopRep,
+                                BlockTSmallerVector &ExitingBlocks,
+                                BlockTSmallerVector &ExitBlocks);
+  BlockT *normalizeInfiniteLoopExit(LoopT *LoopRep);
+  void removeUnconditionalBranch(BlockT *SrcBlock);
+  void removeRedundantConditionalBranch(BlockT *SrcBlock);
+  void addDummyExitBlock(SmallVector<BlockT *, DEFAULT_VEC_SLOTS> &RetBlocks);
+
+  void removeSuccessor(BlockT *SrcBlock);
+  BlockT *cloneBlockForPredecessor(BlockT *CurBlock, BlockT *PredBlock);
+  BlockT *exitingBlock2ExitBlock (LoopT *LoopRep, BlockT *exitingBlock);
+
+  void migrateInstruction(BlockT *SrcBlock, BlockT *DstBlock,
+                          InstrIterator InsertPos);
+
+  void recordSccnum(BlockT *SrcBlock, int SCCNum);
+  int getSCCNum(BlockT *srcBlk);
+
+  void retireBlock(BlockT *DstBlock, BlockT *SrcBlock);
+  bool isRetiredBlock(BlockT *SrcBlock);
+  bool isActiveLoophead(BlockT *CurBlock);
+  bool needMigrateBlock(BlockT *Block);
+
+  BlockT *recordLoopLandBlock(LoopT *LoopRep, BlockT *LandBlock,
+                              BlockTSmallerVector &exitBlocks,
+                              std::set<BlockT*> &ExitBlockSet);
+  void setLoopLandBlock(LoopT *LoopRep, BlockT *Block = NULL);
+  BlockT *getLoopLandBlock(LoopT *LoopRep);
+  LoopLandInfo *getLoopLandInfo(LoopT *LoopRep);
+
+  void addLoopBreakOnReg(LoopT *LoopRep, RegiT RegNum);
+  void addLoopContOnReg(LoopT *LoopRep, RegiT RegNum);
+  void addLoopBreakInitReg(LoopT *LoopRep, RegiT RegNum);
+  void addLoopContInitReg(LoopT *LoopRep, RegiT RegNum);
+  void addLoopEndbranchInitReg(LoopT *LoopRep, RegiT RegNum);
+
+  bool hasBackEdge(BlockT *curBlock);
+  unsigned getLoopDepth  (LoopT *LoopRep);
+  int countActiveBlock(
+    typename SmallVector<BlockT *, DEFAULT_VEC_SLOTS>::const_iterator IterStart,
+    typename SmallVector<BlockT *, DEFAULT_VEC_SLOTS>::const_iterator IterEnd);
+    BlockT *findNearestCommonPostDom(std::set<BlockT *>&);
+  BlockT *findNearestCommonPostDom(BlockT *Block1, BlockT *Block2);
+
+private:
+  DomTreeT *domTree;
+  PostDomTreeT *postDomTree;
+  LoopInfoT *loopInfo;
+  PassT *passRep;
+  FuncT *funcRep;
+
+  BlockInfoMap blockInfoMap;
+  LoopLandInfoMap loopLandInfoMap;
+  SmallVector<BlockT *, DEFAULT_VEC_SLOTS> orderedBlks;
+  const AMDGPURegisterInfo *TRI;
+
+};  //template class CFGStructurizer
+
+template<class PassT> CFGStructurizer<PassT>::CFGStructurizer()
+  : domTree(NULL), postDomTree(NULL), loopInfo(NULL) {
+}
+
+template<class PassT> CFGStructurizer<PassT>::~CFGStructurizer() {
+  for (typename BlockInfoMap::iterator I = blockInfoMap.begin(),
+       E = blockInfoMap.end(); I != E; ++I) {
+    delete I->second;
+  }
+}
+
+template<class PassT>
+bool CFGStructurizer<PassT>::prepare(FuncT &func, PassT &pass,
+                                     const AMDGPURegisterInfo * tri) {
+  passRep = &pass;
+  funcRep = &func;
+  TRI = tri;
+
+  bool changed = false;
+
+  //FIXME: if not reducible flow graph, make it so ???
+
+  if (DEBUGME) {
+        errs() << "AMDGPUCFGStructurizer::prepare\n";
+  }
+
+  loopInfo = CFGTraits::getLoopInfo(pass);
+  if (DEBUGME) {
+    errs() << "LoopInfo:\n";
+    PrintLoopinfo(*loopInfo, errs());
+  }
+
+  orderBlocks();
+  if (DEBUGME) {
+    errs() << "Ordered blocks:\n";
+    printOrderedBlocks(errs());
+  }
+
+  SmallVector<BlockT *, DEFAULT_VEC_SLOTS> retBlks;
+
+  for (typename LoopInfoT::iterator iter = loopInfo->begin(),
+       iterEnd = loopInfo->end();
+       iter != iterEnd; ++iter) {
+    LoopT* loopRep = (*iter);
+    BlockTSmallerVector exitingBlks;
+    loopRep->getExitingBlocks(exitingBlks);
+    
+    if (exitingBlks.size() == 0) {
+      BlockT* dummyExitBlk = normalizeInfiniteLoopExit(loopRep);
+      if (dummyExitBlk != NULL)
+        retBlks.push_back(dummyExitBlk);
+    }
+  }
+
+  // Remove unconditional branch instr.
+  // Add dummy exit block iff there are multiple returns.
+
+  for (typename SmallVector<BlockT *, DEFAULT_VEC_SLOTS>::const_iterator
+       iterBlk = orderedBlks.begin(), iterEndBlk = orderedBlks.end();
+       iterBlk != iterEndBlk;
+       ++iterBlk) {
+    BlockT *curBlk = *iterBlk;
+    removeUnconditionalBranch(curBlk);
+    removeRedundantConditionalBranch(curBlk);
+    if (CFGTraits::isReturnBlock(curBlk)) {
+      retBlks.push_back(curBlk);
+    }
+    assert(curBlk->succ_size() <= 2);
+  } //for
+
+  if (retBlks.size() >= 2) {
+    addDummyExitBlock(retBlks);
+    changed = true;
+  }
+
+  return changed;
+} //CFGStructurizer::prepare
+
+template<class PassT>
+bool CFGStructurizer<PassT>::run(FuncT &func, PassT &pass,
+    const AMDGPURegisterInfo * tri) {
+  passRep = &pass;
+  funcRep = &func;
+  TRI = tri;
+
+  //Assume reducible CFG...
+  if (DEBUGME) {
+    errs() << "AMDGPUCFGStructurizer::run\n";
+    func.viewCFG();
+  }
+
+  domTree = CFGTraits::getDominatorTree(pass);
+  if (DEBUGME) {
+    domTree->print(errs(), (const llvm::Module*)0);
+  }
+
+  postDomTree = CFGTraits::getPostDominatorTree(pass);
+  if (DEBUGME) {
+    postDomTree->print(errs());
+  }
+
+  loopInfo = CFGTraits::getLoopInfo(pass);
+  if (DEBUGME) {
+    errs() << "LoopInfo:\n";
+    PrintLoopinfo(*loopInfo, errs());
+  }
+
+  orderBlocks();
+#ifdef STRESSTEST
+  //Use the worse block ordering to test the algorithm.
+  ReverseVector(orderedBlks);
+#endif
+
+  if (DEBUGME) {
+    errs() << "Ordered blocks:\n";
+    printOrderedBlocks(errs());
+  }
+  int numIter = 0;
+  bool finish = false;
+  BlockT *curBlk;
+  bool makeProgress = false;
+  int numRemainedBlk = countActiveBlock(orderedBlks.begin(),
+                                        orderedBlks.end());
+
+  do {
+    ++numIter;
+    if (DEBUGME) {
+      errs() << "numIter = " << numIter
+             << ", numRemaintedBlk = " << numRemainedBlk << "\n";
+    }
+
+    typename SmallVector<BlockT *, DEFAULT_VEC_SLOTS>::const_iterator
+      iterBlk = orderedBlks.begin();
+    typename SmallVector<BlockT *, DEFAULT_VEC_SLOTS>::const_iterator
+      iterBlkEnd = orderedBlks.end();
+
+    typename SmallVector<BlockT *, DEFAULT_VEC_SLOTS>::const_iterator
+      sccBeginIter = iterBlk;
+    BlockT *sccBeginBlk = NULL;
+    int sccNumBlk = 0;  // The number of active blocks, init to a
+                        // maximum possible number.
+    int sccNumIter;     // Number of iteration in this SCC.
+
+    while (iterBlk != iterBlkEnd) {
+      curBlk = *iterBlk;
+
+      if (sccBeginBlk == NULL) {
+        sccBeginIter = iterBlk;
+        sccBeginBlk = curBlk;
+        sccNumIter = 0;
+        sccNumBlk = numRemainedBlk; // Init to maximum possible number.
+        if (DEBUGME) {
+              errs() << "start processing SCC" << getSCCNum(sccBeginBlk);
+              errs() << "\n";
+        }
+      }
+
+      if (!isRetiredBlock(curBlk)) {
+        patternMatch(curBlk);
+      }
+
+      ++iterBlk;
+
+      bool contNextScc = true;
+      if (iterBlk == iterBlkEnd
+          || getSCCNum(sccBeginBlk) != getSCCNum(*iterBlk)) {
+        // Just finish one scc.
+        ++sccNumIter;
+        int sccRemainedNumBlk = countActiveBlock(sccBeginIter, iterBlk);
+        if (sccRemainedNumBlk != 1 && sccRemainedNumBlk >= sccNumBlk) {
+          if (DEBUGME) {
+            errs() << "Can't reduce SCC " << getSCCNum(curBlk)
+                   << ", sccNumIter = " << sccNumIter;
+            errs() << "doesn't make any progress\n";
+          }
+          contNextScc = true;
+        } else if (sccRemainedNumBlk != 1 && sccRemainedNumBlk < sccNumBlk) {
+          sccNumBlk = sccRemainedNumBlk;
+          iterBlk = sccBeginIter;
+          contNextScc = false;
+          if (DEBUGME) {
+            errs() << "repeat processing SCC" << getSCCNum(curBlk)
+                   << "sccNumIter = " << sccNumIter << "\n";
+            func.viewCFG();
+          }
+        } else {
+          // Finish the current scc.
+          contNextScc = true;
+        }
+      } else {
+        // Continue on next component in the current scc.
+        contNextScc = false;
+      }
+
+      if (contNextScc) {
+        sccBeginBlk = NULL;
+      }
+    } //while, "one iteration" over the function.
+
+    BlockT *entryBlk = FuncGTraits::nodes_begin(&func);
+    if (entryBlk->succ_size() == 0) {
+      finish = true;
+      if (DEBUGME) {
+        errs() << "Reduce to one block\n";
+      }
+    } else {
+      int newnumRemainedBlk
+        = countActiveBlock(orderedBlks.begin(), orderedBlks.end());
+      // consider cloned blocks ??
+      if (newnumRemainedBlk == 1 || newnumRemainedBlk < numRemainedBlk) {
+        makeProgress = true;
+        numRemainedBlk = newnumRemainedBlk;
+      } else {
+        makeProgress = false;
+        if (DEBUGME) {
+          errs() << "No progress\n";
+        }
+      }
+    }
+  } while (!finish && makeProgress);
+
+  // Misc wrap up to maintain the consistency of the Function representation.
+  CFGTraits::wrapup(FuncGTraits::nodes_begin(&func));
+
+  // Detach retired Block, release memory.
+  for (typename BlockInfoMap::iterator iterMap = blockInfoMap.begin(),
+       iterEndMap = blockInfoMap.end(); iterMap != iterEndMap; ++iterMap) {
+    if ((*iterMap).second && (*iterMap).second->isRetired) {
+      assert(((*iterMap).first)->getNumber() != -1);
+      if (DEBUGME) {
+        errs() << "Erase BB" << ((*iterMap).first)->getNumber() << "\n";
+      }
+      (*iterMap).first->eraseFromParent();  //Remove from the parent Function.
+    }
+    delete (*iterMap).second;
+  }
+  blockInfoMap.clear();
+
+  // clear loopLandInfoMap
+  for (typename LoopLandInfoMap::iterator iterMap = loopLandInfoMap.begin(),
+       iterEndMap = loopLandInfoMap.end(); iterMap != iterEndMap; ++iterMap) {
+    delete (*iterMap).second;
+  }
+  loopLandInfoMap.clear();
+
+  if (DEBUGME) {
+    func.viewCFG();
+  }
+
+  if (!finish) {
+    assert(!"IRREDUCIBL_CF");
+  }
+
+  return true;
+} //CFGStructurizer::run
+
+/// Print the ordered Blocks.
+///
+template<class PassT>
+void CFGStructurizer<PassT>::printOrderedBlocks(llvm::raw_ostream &os) {
+  size_t i = 0;
+  for (typename SmallVector<BlockT *, DEFAULT_VEC_SLOTS>::const_iterator
+      iterBlk = orderedBlks.begin(), iterBlkEnd = orderedBlks.end();
+       iterBlk != iterBlkEnd;
+       ++iterBlk, ++i) {
+    os << "BB" << (*iterBlk)->getNumber();
+    os << "(" << getSCCNum(*iterBlk) << "," << (*iterBlk)->size() << ")";
+    if (i != 0 && i % 10 == 0) {
+      os << "\n";
+    } else {
+      os << " ";
+    }
+  }
+} //printOrderedBlocks
+
+/// Compute the reversed DFS post order of Blocks
+///
+template<class PassT> void CFGStructurizer<PassT>::orderBlocks() {
+  int sccNum = 0;
+  BlockT *bb;
+  for (scc_iterator<FuncT *> sccIter = scc_begin(funcRep),
+       sccEnd = scc_end(funcRep); sccIter != sccEnd; ++sccIter, ++sccNum) {
+    std::vector<BlockT *> &sccNext = *sccIter;
+    for (typename std::vector<BlockT *>::const_iterator
+         blockIter = sccNext.begin(), blockEnd = sccNext.end();
+         blockIter != blockEnd; ++blockIter) {
+      bb = *blockIter;
+      orderedBlks.push_back(bb);
+      recordSccnum(bb, sccNum);
+    }
+  }
+
+  //walk through all the block in func to check for unreachable
+  for (BlockIterator blockIter1 = FuncGTraits::nodes_begin(funcRep),
+       blockEnd1 = FuncGTraits::nodes_end(funcRep);
+       blockIter1 != blockEnd1; ++blockIter1) {
+    BlockT *bb = &(*blockIter1);
+    sccNum = getSCCNum(bb);
+    if (sccNum == INVALIDSCCNUM) {
+      errs() << "unreachable block BB" << bb->getNumber() << "\n";
+    }
+  }
+} //orderBlocks
+
+template<class PassT> int CFGStructurizer<PassT>::patternMatch(BlockT *curBlk) {
+  int numMatch = 0;
+  int curMatch;
+
+  if (DEBUGME) {
+        errs() << "Begin patternMatch BB" << curBlk->getNumber() << "\n";
+  }
+
+  while ((curMatch = patternMatchGroup(curBlk)) > 0) {
+    numMatch += curMatch;
+  }
+
+  if (DEBUGME) {
+        errs() << "End patternMatch BB" << curBlk->getNumber()
+      << ", numMatch = " << numMatch << "\n";
+  }
+
+  return numMatch;
+} //patternMatch
+
+template<class PassT>
+int CFGStructurizer<PassT>::patternMatchGroup(BlockT *curBlk) {
+  int numMatch = 0;
+  numMatch += serialPatternMatch(curBlk);
+  numMatch += ifPatternMatch(curBlk);
+  numMatch += loopendPatternMatch(curBlk);
+  numMatch += loopPatternMatch(curBlk);
+  return numMatch;
+}//patternMatchGroup
+
+template<class PassT>
+int CFGStructurizer<PassT>::serialPatternMatch(BlockT *curBlk) {
+  if (curBlk->succ_size() != 1) {
+    return 0;
+  }
+
+  BlockT *childBlk = *curBlk->succ_begin();
+  if (childBlk->pred_size() != 1 || isActiveLoophead(childBlk)) {
+    return 0;
+  }
+
+  mergeSerialBlock(curBlk, childBlk);
+  ++numSerialPatternMatch;
+  return 1;
+} //serialPatternMatch
+
+template<class PassT>
+int CFGStructurizer<PassT>::ifPatternMatch(BlockT *curBlk) {
+  //two edges
+  if (curBlk->succ_size() != 2) {
+    return 0;
+  }
+
+  if (hasBackEdge(curBlk)) {
+    return 0;
+  }
+
+  InstrT *branchInstr = CFGTraits::getNormalBlockBranchInstr(curBlk);
+  if (branchInstr == NULL) {
+    return 0;
+  }
+
+  assert(CFGTraits::isCondBranch(branchInstr));
+
+  BlockT *trueBlk = CFGTraits::getTrueBranch(branchInstr);
+  BlockT *falseBlk = CFGTraits::getFalseBranch(curBlk, branchInstr);
+  BlockT *landBlk;
+  int cloned = 0;
+
+  // TODO: Simplify
+  if (trueBlk->succ_size() == 1 && falseBlk->succ_size() == 1
+    && *trueBlk->succ_begin() == *falseBlk->succ_begin()) {
+    landBlk = *trueBlk->succ_begin();
+  } else if (trueBlk->succ_size() == 0 && falseBlk->succ_size() == 0) {
+    landBlk = NULL;
+  } else if (trueBlk->succ_size() == 1 && *trueBlk->succ_begin() == falseBlk) {
+    landBlk = falseBlk;
+    falseBlk = NULL;
+  } else if (falseBlk->succ_size() == 1
+             && *falseBlk->succ_begin() == trueBlk) {
+    landBlk = trueBlk;
+    trueBlk = NULL;
+  } else if (falseBlk->succ_size() == 1
+             && isSameloopDetachedContbreak(trueBlk, falseBlk)) {
+    landBlk = *falseBlk->succ_begin();
+  } else if (trueBlk->succ_size() == 1
+    && isSameloopDetachedContbreak(falseBlk, trueBlk)) {
+    landBlk = *trueBlk->succ_begin();
+  } else {
+    return handleJumpintoIf(curBlk, trueBlk, falseBlk);
+  }
+
+  // improveSimpleJumpinfoIf can handle the case where landBlk == NULL but the
+  // new BB created for landBlk==NULL may introduce new challenge to the
+  // reduction process.
+  if (landBlk != NULL &&
+      ((trueBlk && trueBlk->pred_size() > 1)
+      || (falseBlk && falseBlk->pred_size() > 1))) {
+     cloned += improveSimpleJumpintoIf(curBlk, trueBlk, falseBlk, &landBlk);
+  }
+
+  if (trueBlk && trueBlk->pred_size() > 1) {
+    trueBlk = cloneBlockForPredecessor(trueBlk, curBlk);
+    ++cloned;
+  }
+
+  if (falseBlk && falseBlk->pred_size() > 1) {
+    falseBlk = cloneBlockForPredecessor(falseBlk, curBlk);
+    ++cloned;
+  }
+
+  mergeIfthenelseBlock(branchInstr, curBlk, trueBlk, falseBlk, landBlk);
+
+  ++numIfPatternMatch;
+
+  numClonedBlock += cloned;
+
+  return 1 + cloned;
+} //ifPatternMatch
+
+template<class PassT>
+int CFGStructurizer<PassT>::switchPatternMatch(BlockT *curBlk) {
+  return 0;
+} //switchPatternMatch
+
+template<class PassT>
+int CFGStructurizer<PassT>::loopendPatternMatch(BlockT *curBlk) {
+  LoopT *loopRep = loopInfo->getLoopFor(curBlk);
+  typename std::vector<LoopT *> nestedLoops;
+  while (loopRep) {
+    nestedLoops.push_back(loopRep);
+    loopRep = loopRep->getParentLoop();
+  }
+
+  if (nestedLoops.size() == 0) {
+    return 0;
+  }
+
+  // Process nested loop outside->inside, so "continue" to a outside loop won't
+  // be mistaken as "break" of the current loop.
+  int num = 0;
+  for (typename std::vector<LoopT *>::reverse_iterator
+       iter = nestedLoops.rbegin(), iterEnd = nestedLoops.rend();
+       iter != iterEnd; ++iter) {
+    loopRep = *iter;
+
+    if (getLoopLandBlock(loopRep) != NULL) {
+      continue;
+    }
+
+    BlockT *loopHeader = loopRep->getHeader();
+
+    int numBreak = loopbreakPatternMatch(loopRep, loopHeader);
+
+    if (numBreak == -1) {
+      break;
+    }
+
+    int numCont = loopcontPatternMatch(loopRep, loopHeader);
+    num += numBreak + numCont;
+  }
+
+  return num;
+} //loopendPatternMatch
+
+template<class PassT>
+int CFGStructurizer<PassT>::loopPatternMatch(BlockT *curBlk) {
+  if (curBlk->succ_size() != 0) {
+    return 0;
+  }
+
+  int numLoop = 0;
+  LoopT *loopRep = loopInfo->getLoopFor(curBlk);
+  while (loopRep && loopRep->getHeader() == curBlk) {
+    LoopLandInfo *loopLand = getLoopLandInfo(loopRep);
+    if (loopLand) {
+      BlockT *landBlk = loopLand->landBlk;
+      assert(landBlk);
+      if (!isRetiredBlock(landBlk)) {
+        mergeLooplandBlock(curBlk, loopLand);
+        ++numLoop;
+      }
+    }
+    loopRep = loopRep->getParentLoop();
+  }
+
+  numLoopPatternMatch += numLoop;
+
+  return numLoop;
+} //loopPatternMatch
+
+template<class PassT>
+int CFGStructurizer<PassT>::loopbreakPatternMatch(LoopT *loopRep,
+                                                  BlockT *loopHeader) {
+  BlockTSmallerVector exitingBlks;
+  loopRep->getExitingBlocks(exitingBlks);
+
+  if (DEBUGME) {
+    errs() << "Loop has " << exitingBlks.size() << " exiting blocks\n";
+  }
+
+  if (exitingBlks.size() == 0) {
+    setLoopLandBlock(loopRep);
+    return 0;
+  }
+
+  // Compute the corresponding exitBlks and exit block set.
+  BlockTSmallerVector exitBlks;
+  std::set<BlockT *> exitBlkSet;
+  for (typename BlockTSmallerVector::const_iterator iter = exitingBlks.begin(),
+       iterEnd = exitingBlks.end(); iter != iterEnd; ++iter) {
+    BlockT *exitingBlk = *iter;
+    BlockT *exitBlk = exitingBlock2ExitBlock(loopRep, exitingBlk);
+    exitBlks.push_back(exitBlk);
+    exitBlkSet.insert(exitBlk);  //non-duplicate insert
+  }
+
+  assert(exitBlkSet.size() > 0);
+  assert(exitBlks.size() == exitingBlks.size());
+
+  if (DEBUGME) {
+    errs() << "Loop has " << exitBlkSet.size() << " exit blocks\n";
+  }
+
+  // Find exitLandBlk.
+  BlockT *exitLandBlk = NULL;
+  int numCloned = 0;
+  int numSerial = 0;
+
+  if (exitBlkSet.size() == 1) {
+    exitLandBlk = *exitBlkSet.begin();
+  } else {
+    exitLandBlk = findNearestCommonPostDom(exitBlkSet);
+
+    if (exitLandBlk == NULL) {
+      return -1;
+    }
+
+    bool allInPath = true;
+    bool allNotInPath = true;
+    for (typename std::set<BlockT*>::const_iterator
+         iter = exitBlkSet.begin(),
+         iterEnd = exitBlkSet.end();
+         iter != iterEnd; ++iter) {
+      BlockT *exitBlk = *iter;
+
+      PathToKind pathKind = singlePathTo(exitBlk, exitLandBlk, true);
+      if (DEBUGME) {
+        errs() << "BB" << exitBlk->getNumber()
+               << " to BB" << exitLandBlk->getNumber() << " PathToKind="
+               << pathKind << "\n";
+      }
+
+      allInPath = allInPath && (pathKind == SinglePath_InPath);
+      allNotInPath = allNotInPath && (pathKind == SinglePath_NotInPath);
+
+      if (!allInPath && !allNotInPath) {
+        if (DEBUGME) {
+              errs() << "singlePath check fail\n";
+        }
+        return -1;
+      }
+    } // check all exit blocks
+
+    if (allNotInPath) {
+
+      // TODO: Simplify, maybe separate function?
+      LoopT *parentLoopRep = loopRep->getParentLoop();
+      BlockT *parentLoopHeader = NULL;
+      if (parentLoopRep)
+        parentLoopHeader = parentLoopRep->getHeader();
+
+      if (exitLandBlk == parentLoopHeader &&
+          (exitLandBlk = relocateLoopcontBlock(parentLoopRep,
+                                               loopRep,
+                                               exitBlkSet,
+                                               exitLandBlk)) != NULL) {
+        if (DEBUGME) {
+          errs() << "relocateLoopcontBlock success\n";
+        }
+      } else if ((exitLandBlk = addLoopEndbranchBlock(loopRep,
+                                                      exitingBlks,
+                                                      exitBlks)) != NULL) {
+        if (DEBUGME) {
+          errs() << "insertEndbranchBlock success\n";
+        }
+      } else {
+        if (DEBUGME) {
+          errs() << "loop exit fail\n";
+        }
+        return -1;
+      }
+    }
+
+    // Handle side entry to exit path.
+    exitBlks.clear();
+    exitBlkSet.clear();
+    for (typename BlockTSmallerVector::iterator iterExiting =
+           exitingBlks.begin(),
+         iterExitingEnd = exitingBlks.end();
+         iterExiting != iterExitingEnd; ++iterExiting) {
+      BlockT *exitingBlk = *iterExiting;
+      BlockT *exitBlk = exitingBlock2ExitBlock(loopRep, exitingBlk);
+      BlockT *newExitBlk = exitBlk;
+
+      if (exitBlk != exitLandBlk && exitBlk->pred_size() > 1) {
+        newExitBlk = cloneBlockForPredecessor(exitBlk, exitingBlk);
+        ++numCloned;
+      }
+
+      numCloned += cloneOnSideEntryTo(exitingBlk, newExitBlk, exitLandBlk);
+
+      exitBlks.push_back(newExitBlk);
+      exitBlkSet.insert(newExitBlk);
+    }
+
+    for (typename BlockTSmallerVector::iterator iterExit = exitBlks.begin(),
+         iterExitEnd = exitBlks.end();
+         iterExit != iterExitEnd; ++iterExit) {
+      BlockT *exitBlk = *iterExit;
+      numSerial += serialPatternMatch(exitBlk);
+    }
+
+    for (typename BlockTSmallerVector::iterator iterExit = exitBlks.begin(),
+         iterExitEnd = exitBlks.end();
+         iterExit != iterExitEnd; ++iterExit) {
+      BlockT *exitBlk = *iterExit;
+      if (exitBlk->pred_size() > 1) {
+        if (exitBlk != exitLandBlk) {
+          return -1;
+        }
+      } else {
+        if (exitBlk != exitLandBlk &&
+            (exitBlk->succ_size() != 1 ||
+            *exitBlk->succ_begin() != exitLandBlk)) {
+          return -1;
+        }
+      }
+    }
+  } // else
+
+  exitLandBlk = recordLoopLandBlock(loopRep, exitLandBlk, exitBlks, exitBlkSet);
+
+  // Fold break into the breaking block. Leverage across level breaks.
+  assert(exitingBlks.size() == exitBlks.size());
+  for (typename BlockTSmallerVector::const_iterator iterExit = exitBlks.begin(),
+       iterExiting = exitingBlks.begin(), iterExitEnd = exitBlks.end();
+       iterExit != iterExitEnd; ++iterExit, ++iterExiting) {
+    BlockT *exitBlk = *iterExit;
+    BlockT *exitingBlk = *iterExiting;
+    assert(exitBlk->pred_size() == 1 || exitBlk == exitLandBlk);
+    LoopT *exitingLoop = loopInfo->getLoopFor(exitingBlk);
+    handleLoopbreak(exitingBlk, exitingLoop, exitBlk, loopRep, exitLandBlk);
+  }
+
+  int numBreak = static_cast<int>(exitingBlks.size());
+  numLoopbreakPatternMatch += numBreak;
+  numClonedBlock += numCloned;
+  return numBreak + numSerial + numCloned;
+} //loopbreakPatternMatch
+
+template<class PassT>
+int CFGStructurizer<PassT>::loopcontPatternMatch(LoopT *loopRep,
+                                                 BlockT *loopHeader) {
+  int numCont = 0;
+  SmallVector<BlockT *, DEFAULT_VEC_SLOTS> contBlk;
+  for (typename InvBlockGTraits::ChildIteratorType iter =
+       InvBlockGTraits::child_begin(loopHeader),
+       iterEnd = InvBlockGTraits::child_end(loopHeader);
+       iter != iterEnd; ++iter) {
+    BlockT *curBlk = *iter;
+    if (loopRep->contains(curBlk)) {
+      handleLoopcontBlock(curBlk, loopInfo->getLoopFor(curBlk),
+                          loopHeader, loopRep);
+      contBlk.push_back(curBlk);
+      ++numCont;
+    }
+  }
+
+  for (typename SmallVector<BlockT *, DEFAULT_VEC_SLOTS>::iterator
+       iter = contBlk.begin(), iterEnd = contBlk.end();
+       iter != iterEnd; ++iter) {
+    (*iter)->removeSuccessor(loopHeader);
+  }
+
+  numLoopcontPatternMatch += numCont;
+
+  return numCont;
+} //loopcontPatternMatch
+
+
+template<class PassT>
+bool CFGStructurizer<PassT>::isSameloopDetachedContbreak(BlockT *src1Blk,
+                                                         BlockT *src2Blk) {
+  // return true iff src1Blk->succ_size() == 0 && src1Blk and src2Blk are in the
+  // same loop with LoopLandInfo without explicitly keeping track of
+  // loopContBlks and loopBreakBlks, this is a method to get the information.
+  //
+  if (src1Blk->succ_size() == 0) {
+    LoopT *loopRep = loopInfo->getLoopFor(src1Blk);
+    if (loopRep != NULL && loopRep == loopInfo->getLoopFor(src2Blk)) {
+      LoopLandInfo *&theEntry = loopLandInfoMap[loopRep];
+      if (theEntry != NULL) {
+        if (DEBUGME) {
+          errs() << "isLoopContBreakBlock yes src1 = BB"
+                 << src1Blk->getNumber()
+                 << " src2 = BB" << src2Blk->getNumber() << "\n";
+        }
+        return true;
+      }
+    }
+  }
+  return false;
+}  //isSameloopDetachedContbreak
+
+template<class PassT>
+int CFGStructurizer<PassT>::handleJumpintoIf(BlockT *headBlk,
+                                             BlockT *trueBlk,
+                                             BlockT *falseBlk) {
+  int num = handleJumpintoIfImp(headBlk, trueBlk, falseBlk);
+  if (num == 0) {
+    if (DEBUGME) {
+      errs() << "handleJumpintoIf swap trueBlk and FalseBlk" << "\n";
+    }
+    num = handleJumpintoIfImp(headBlk, falseBlk, trueBlk);
+  }
+  return num;
+}
+
+template<class PassT>
+int CFGStructurizer<PassT>::handleJumpintoIfImp(BlockT *headBlk,
+                                                BlockT *trueBlk,
+                                                BlockT *falseBlk) {
+  int num = 0;
+  BlockT *downBlk;
+
+  //trueBlk could be the common post dominator
+  downBlk = trueBlk;
+
+  if (DEBUGME) {
+    errs() << "handleJumpintoIfImp head = BB" << headBlk->getNumber()
+           << " true = BB" << trueBlk->getNumber()
+           << ", numSucc=" << trueBlk->succ_size()
+           << " false = BB" << falseBlk->getNumber() << "\n";
+  }
+
+  while (downBlk) {
+    if (DEBUGME) {
+      errs() << "check down = BB" << downBlk->getNumber();
+    }
+
+    if (singlePathTo(falseBlk, downBlk) == SinglePath_InPath) {
+      if (DEBUGME) {
+        errs() << " working\n";
+      }
+
+      num += cloneOnSideEntryTo(headBlk, trueBlk, downBlk);
+      num += cloneOnSideEntryTo(headBlk, falseBlk, downBlk);
+
+      numClonedBlock += num;
+      num += serialPatternMatch(*headBlk->succ_begin());
+      num += serialPatternMatch(*(++headBlk->succ_begin()));
+      num += ifPatternMatch(headBlk);
+      assert(num > 0);
+
+      break;
+    }
+    if (DEBUGME) {
+      errs() << " not working\n";
+    }
+    downBlk = (downBlk->succ_size() == 1) ? (*downBlk->succ_begin()) : NULL;
+  } // walk down the postDomTree
+
+  return num;
+} //handleJumpintoIf
+
+template<class PassT>
+void CFGStructurizer<PassT>::showImproveSimpleJumpintoIf(BlockT *headBlk,
+                                                         BlockT *trueBlk,
+                                                         BlockT *falseBlk,
+                                                         BlockT *landBlk,
+                                                         bool detail) {
+  errs() << "head = BB" << headBlk->getNumber()
+         << " size = " << headBlk->size();
+  if (detail) {
+    errs() << "\n";
+    headBlk->print(errs());
+    errs() << "\n";
+  }
+
+  if (trueBlk) {
+    errs() << ", true = BB" << trueBlk->getNumber() << " size = "
+           << trueBlk->size() << " numPred = " << trueBlk->pred_size();
+    if (detail) {
+      errs() << "\n";
+      trueBlk->print(errs());
+      errs() << "\n";
+    }
+  }
+  if (falseBlk) {
+    errs() << ", false = BB" << falseBlk->getNumber() << " size = "
+           << falseBlk->size() << " numPred = " << falseBlk->pred_size();
+    if (detail) {
+      errs() << "\n";
+      falseBlk->print(errs());
+      errs() << "\n";
+    }
+  }
+  if (landBlk) {
+    errs() << ", land = BB" << landBlk->getNumber() << " size = "
+           << landBlk->size() << " numPred = " << landBlk->pred_size();
+    if (detail) {
+      errs() << "\n";
+      landBlk->print(errs());
+      errs() << "\n";
+    }
+  }
+
+    errs() << "\n";
+} //showImproveSimpleJumpintoIf
+
+template<class PassT>
+int CFGStructurizer<PassT>::improveSimpleJumpintoIf(BlockT *headBlk,
+                                                    BlockT *trueBlk,
+                                                    BlockT *falseBlk,
+                                                    BlockT **plandBlk) {
+  bool migrateTrue = false;
+  bool migrateFalse = false;
+
+  BlockT *landBlk = *plandBlk;
+
+  assert((trueBlk == NULL || trueBlk->succ_size() <= 1)
+         && (falseBlk == NULL || falseBlk->succ_size() <= 1));
+
+  if (trueBlk == falseBlk) {
+    return 0;
+  }
+
+  migrateTrue = needMigrateBlock(trueBlk);
+  migrateFalse = needMigrateBlock(falseBlk);
+
+  if (!migrateTrue && !migrateFalse) {
+    return 0;
+  }
+
+  // If we need to migrate either trueBlk and falseBlk, migrate the rest that
+  // have more than one predecessors.  without doing this, its predecessor
+  // rather than headBlk will have undefined value in initReg.
+  if (!migrateTrue && trueBlk && trueBlk->pred_size() > 1) {
+    migrateTrue = true;
+  }
+  if (!migrateFalse && falseBlk && falseBlk->pred_size() > 1) {
+    migrateFalse = true;
+  }
+
+  if (DEBUGME) {
+    errs() << "before improveSimpleJumpintoIf: ";
+    showImproveSimpleJumpintoIf(headBlk, trueBlk, falseBlk, landBlk, 0);
+  }
+
+  // org: headBlk => if () {trueBlk} else {falseBlk} => landBlk
+  //
+  // new: headBlk => if () {initReg = 1; org trueBlk branch} else
+  //      {initReg = 0; org falseBlk branch }
+  //      => landBlk => if (initReg) {org trueBlk} else {org falseBlk}
+  //      => org landBlk
+  //      if landBlk->pred_size() > 2, put the about if-else inside
+  //      if (initReg !=2) {...}
+  //
+  // add initReg = initVal to headBlk
+
+  const TargetRegisterClass * I32RC = TRI->getCFGStructurizerRegClass(MVT::i32);
+  unsigned initReg =
+    funcRep->getRegInfo().createVirtualRegister(I32RC);
+  if (!migrateTrue || !migrateFalse) {
+    int initVal = migrateTrue ? 0 : 1;
+    CFGTraits::insertAssignInstrBefore(headBlk, passRep, initReg, initVal);
+  }
+
+  int numNewBlk = 0;
+
+  if (landBlk == NULL) {
+    landBlk = funcRep->CreateMachineBasicBlock();
+    funcRep->push_back(landBlk);  //insert to function
+
+    if (trueBlk) {
+      trueBlk->addSuccessor(landBlk);
+    } else {
+      headBlk->addSuccessor(landBlk);
+    }
+
+    if (falseBlk) {
+      falseBlk->addSuccessor(landBlk);
+    } else {
+      headBlk->addSuccessor(landBlk);
+    }
+
+    numNewBlk ++;
+  }
+
+  bool landBlkHasOtherPred = (landBlk->pred_size() > 2);
+
+  //insert AMDGPU::ENDIF to avoid special case "input landBlk == NULL"
+  typename BlockT::iterator insertPos =
+    CFGTraits::getInstrPos
+    (landBlk, CFGTraits::insertInstrBefore(landBlk, AMDGPU::ENDIF, passRep));
+
+  if (landBlkHasOtherPred) {
+    unsigned immReg =
+      funcRep->getRegInfo().createVirtualRegister(I32RC);
+    CFGTraits::insertAssignInstrBefore(insertPos, passRep, immReg, 2);
+    unsigned cmpResReg =
+      funcRep->getRegInfo().createVirtualRegister(I32RC);
+
+    CFGTraits::insertCompareInstrBefore(landBlk, insertPos, passRep, cmpResReg,
+                                        initReg, immReg);
+    CFGTraits::insertCondBranchBefore(landBlk, insertPos,
+                                      AMDGPU::IF_PREDICATE_SET, passRep,
+                                      cmpResReg, DebugLoc());
+  }
+
+  CFGTraits::insertCondBranchBefore(landBlk, insertPos, AMDGPU::IF_PREDICATE_SET,
+                                    passRep, initReg, DebugLoc());
+
+  if (migrateTrue) {
+    migrateInstruction(trueBlk, landBlk, insertPos);
+    // need to uncondionally insert the assignment to ensure a path from its
+    // predecessor rather than headBlk has valid value in initReg if
+    // (initVal != 1).
+    CFGTraits::insertAssignInstrBefore(trueBlk, passRep, initReg, 1);
+  }
+  CFGTraits::insertInstrBefore(insertPos, AMDGPU::ELSE, passRep);
+
+  if (migrateFalse) {
+    migrateInstruction(falseBlk, landBlk, insertPos);
+    // need to uncondionally insert the assignment to ensure a path from its
+    // predecessor rather than headBlk has valid value in initReg if
+    // (initVal != 0)
+    CFGTraits::insertAssignInstrBefore(falseBlk, passRep, initReg, 0);
+  }
+
+  if (landBlkHasOtherPred) {
+    // add endif
+    CFGTraits::insertInstrBefore(insertPos, AMDGPU::ENDIF, passRep);
+
+    // put initReg = 2 to other predecessors of landBlk
+    for (typename BlockT::pred_iterator predIter = landBlk->pred_begin(),
+         predIterEnd = landBlk->pred_end(); predIter != predIterEnd;
+         ++predIter) {
+      BlockT *curBlk = *predIter;
+      if (curBlk != trueBlk && curBlk != falseBlk) {
+        CFGTraits::insertAssignInstrBefore(curBlk, passRep, initReg, 2);
+      }
+    } //for
+  }
+  if (DEBUGME) {
+    errs() << "result from improveSimpleJumpintoIf: ";
+    showImproveSimpleJumpintoIf(headBlk, trueBlk, falseBlk, landBlk, 0);
+  }
+
+  // update landBlk
+  *plandBlk = landBlk;
+
+  return numNewBlk;
+} //improveSimpleJumpintoIf
+
+template<class PassT>
+void CFGStructurizer<PassT>::handleLoopbreak(BlockT *exitingBlk,
+                                              LoopT *exitingLoop,
+                                             BlockT *exitBlk,
+                                              LoopT *exitLoop,
+                                             BlockT *landBlk) {
+  if (DEBUGME) {
+    errs() << "Trying to break loop-depth = " << getLoopDepth(exitLoop)
+           << " from loop-depth = " << getLoopDepth(exitingLoop) << "\n";
+  }
+  const TargetRegisterClass * I32RC = TRI->getCFGStructurizerRegClass(MVT::i32);
+
+  RegiT initReg = INVALIDREGNUM;
+  if (exitingLoop != exitLoop) {
+    initReg = static_cast<int>
+      (funcRep->getRegInfo().createVirtualRegister(I32RC));
+    assert(initReg != INVALIDREGNUM);
+    addLoopBreakInitReg(exitLoop, initReg);
+    while (exitingLoop != exitLoop && exitingLoop) {
+      addLoopBreakOnReg(exitingLoop, initReg);
+      exitingLoop = exitingLoop->getParentLoop();
+    }
+    assert(exitingLoop == exitLoop);
+  }
+
+  mergeLoopbreakBlock(exitingBlk, exitBlk, landBlk, initReg);
+
+} //handleLoopbreak
+
+template<class PassT>
+void CFGStructurizer<PassT>::handleLoopcontBlock(BlockT *contingBlk,
+                                                  LoopT *contingLoop,
+                                                 BlockT *contBlk,
+                                                  LoopT *contLoop) {
+  if (DEBUGME) {
+    errs() << "loopcontPattern cont = BB" << contingBlk->getNumber()
+           << " header = BB" << contBlk->getNumber() << "\n";
+
+    errs() << "Trying to continue loop-depth = "
+           << getLoopDepth(contLoop)
+           << " from loop-depth = " << getLoopDepth(contingLoop) << "\n";
+  }
+
+  RegiT initReg = INVALIDREGNUM;
+  const TargetRegisterClass * I32RC = TRI->getCFGStructurizerRegClass(MVT::i32);
+  if (contingLoop != contLoop) {
+    initReg = static_cast<int>
+      (funcRep->getRegInfo().createVirtualRegister(I32RC));
+    assert(initReg != INVALIDREGNUM);
+    addLoopContInitReg(contLoop, initReg);
+    while (contingLoop && contingLoop->getParentLoop() != contLoop) {
+      addLoopBreakOnReg(contingLoop, initReg);  //not addLoopContOnReg
+      contingLoop = contingLoop->getParentLoop();
+    }
+    assert(contingLoop && contingLoop->getParentLoop() == contLoop);
+    addLoopContOnReg(contingLoop, initReg);
+  }
+
+  settleLoopcontBlock(contingBlk, contBlk, initReg);
+} //handleLoopcontBlock
+
+template<class PassT>
+void CFGStructurizer<PassT>::mergeSerialBlock(BlockT *dstBlk, BlockT *srcBlk) {
+  if (DEBUGME) {
+    errs() << "serialPattern BB" << dstBlk->getNumber()
+           << " <= BB" << srcBlk->getNumber() << "\n";
+  }
+  dstBlk->splice(dstBlk->end(), srcBlk, srcBlk->begin(), srcBlk->end());
+
+  dstBlk->removeSuccessor(srcBlk);
+  CFGTraits::cloneSuccessorList(dstBlk, srcBlk);
+
+  removeSuccessor(srcBlk);
+  retireBlock(dstBlk, srcBlk);
+} //mergeSerialBlock
+
+template<class PassT>
+void CFGStructurizer<PassT>::mergeIfthenelseBlock(InstrT *branchInstr,
+                                                  BlockT *curBlk,
+                                                  BlockT *trueBlk,
+                                                  BlockT *falseBlk,
+                                                  BlockT *landBlk) {
+  if (DEBUGME) {
+    errs() << "ifPattern BB" << curBlk->getNumber();
+    errs() << "{  ";
+    if (trueBlk) {
+      errs() << "BB" << trueBlk->getNumber();
+    }
+    errs() << "  } else ";
+    errs() << "{  ";
+    if (falseBlk) {
+      errs() << "BB" << falseBlk->getNumber();
+    }
+    errs() << "  }\n ";
+    errs() << "landBlock: ";
+    if (landBlk == NULL) {
+      errs() << "NULL";
+    } else {
+      errs() << "BB" << landBlk->getNumber();
+    }
+    errs() << "\n";
+  }
+
+  int oldOpcode = branchInstr->getOpcode();
+  DebugLoc branchDL = branchInstr->getDebugLoc();
+
+//    transform to
+//    if cond
+//       trueBlk
+//    else
+//       falseBlk
+//    endif
+//    landBlk
+
+  typename BlockT::iterator branchInstrPos =
+    CFGTraits::getInstrPos(curBlk, branchInstr);
+  CFGTraits::insertCondBranchBefore(branchInstrPos,
+                                    CFGTraits::getBranchNzeroOpcode(oldOpcode),
+                                    passRep,
+                                    branchDL);
+
+  if (trueBlk) {
+    curBlk->splice(branchInstrPos, trueBlk, trueBlk->begin(), trueBlk->end());
+    curBlk->removeSuccessor(trueBlk);
+    if (landBlk && trueBlk->succ_size()!=0) {
+      trueBlk->removeSuccessor(landBlk);
+    }
+    retireBlock(curBlk, trueBlk);
+  }
+  CFGTraits::insertInstrBefore(branchInstrPos, AMDGPU::ELSE, passRep);
+
+  if (falseBlk) {
+    curBlk->splice(branchInstrPos, falseBlk, falseBlk->begin(),
+                   falseBlk->end());
+    curBlk->removeSuccessor(falseBlk);
+    if (landBlk && falseBlk->succ_size() != 0) {
+      falseBlk->removeSuccessor(landBlk);
+    }
+    retireBlock(curBlk, falseBlk);
+  }
+  CFGTraits::insertInstrBefore(branchInstrPos, AMDGPU::ENDIF, passRep);
+
+  branchInstr->eraseFromParent();
+
+  if (landBlk && trueBlk && falseBlk) {
+    curBlk->addSuccessor(landBlk);
+  }
+
+} //mergeIfthenelseBlock
+
+template<class PassT>
+void CFGStructurizer<PassT>::mergeLooplandBlock(BlockT *dstBlk,
+                                                LoopLandInfo *loopLand) {
+  BlockT *landBlk = loopLand->landBlk;
+
+  if (DEBUGME) {
+    errs() << "loopPattern header = BB" << dstBlk->getNumber()
+           << " land = BB" << landBlk->getNumber() << "\n";
+  }
+
+  // Loop contInitRegs are init at the beginning of the loop.
+  for (typename std::set<RegiT>::const_iterator iter =
+         loopLand->contInitRegs.begin(),
+       iterEnd = loopLand->contInitRegs.end(); iter != iterEnd; ++iter) {
+    CFGTraits::insertAssignInstrBefore(dstBlk, passRep, *iter, 0);
+  }
+
+  /* we last inserterd the DebugLoc in the
+   * BREAK_LOGICALZ_i32 or AMDGPU::BREAK_LOGICALNZ statement in the current dstBlk.
+   * search for the DebugLoc in the that statement.
+   * if not found, we have to insert the empty/default DebugLoc */
+  InstrT *loopBreakInstr = CFGTraits::getLoopBreakInstr(dstBlk);
+  DebugLoc DLBreak = (loopBreakInstr) ? loopBreakInstr->getDebugLoc() : DebugLoc();
+
+  CFGTraits::insertInstrBefore(dstBlk, AMDGPU::WHILELOOP, passRep, DLBreak);
+  // Loop breakInitRegs are init before entering the loop.
+  for (typename std::set<RegiT>::const_iterator iter =
+         loopLand->breakInitRegs.begin(),
+       iterEnd = loopLand->breakInitRegs.end(); iter != iterEnd; ++iter) {
+    CFGTraits::insertAssignInstrBefore(dstBlk, passRep, *iter, 0);
+  }
+  // Loop endbranchInitRegs are init before entering the loop.
+  for (typename std::set<RegiT>::const_iterator iter =
+         loopLand->endbranchInitRegs.begin(),
+       iterEnd = loopLand->endbranchInitRegs.end(); iter != iterEnd; ++iter) {
+    CFGTraits::insertAssignInstrBefore(dstBlk, passRep, *iter, 0);
+  }
+
+  /* we last inserterd the DebugLoc in the continue statement in the current dstBlk
+   * search for the DebugLoc in the continue statement.
+   * if not found, we have to insert the empty/default DebugLoc */
+  InstrT *continueInstr = CFGTraits::getContinueInstr(dstBlk);
+  DebugLoc DLContinue = (continueInstr) ? continueInstr->getDebugLoc() : DebugLoc();
+
+  CFGTraits::insertInstrEnd(dstBlk, AMDGPU::ENDLOOP, passRep, DLContinue);
+  // Loop breakOnRegs are check after the ENDLOOP: break the loop outside this
+  // loop.
+  for (typename std::set<RegiT>::const_iterator iter =
+         loopLand->breakOnRegs.begin(),
+       iterEnd = loopLand->breakOnRegs.end(); iter != iterEnd; ++iter) {
+    CFGTraits::insertCondBranchEnd(dstBlk, AMDGPU::PREDICATED_BREAK, passRep,
+                                   *iter);
+  }
+
+  // Loop contOnRegs are check after the ENDLOOP: cont the loop outside this
+  // loop.
+  for (std::set<RegiT>::const_iterator iter = loopLand->contOnRegs.begin(),
+       iterEnd = loopLand->contOnRegs.end(); iter != iterEnd; ++iter) {
+    CFGTraits::insertCondBranchEnd(dstBlk, AMDGPU::CONTINUE_LOGICALNZ_i32,
+                                   passRep, *iter);
+  }
+
+  dstBlk->splice(dstBlk->end(), landBlk, landBlk->begin(), landBlk->end());
+
+  for (typename BlockT::succ_iterator iter = landBlk->succ_begin(),
+       iterEnd = landBlk->succ_end(); iter != iterEnd; ++iter) {
+    dstBlk->addSuccessor(*iter);  // *iter's predecessor is also taken care of.
+  }
+
+  removeSuccessor(landBlk);
+  retireBlock(dstBlk, landBlk);
+} //mergeLooplandBlock
+
+template<class PassT>
+void CFGStructurizer<PassT>::reversePredicateSetter(typename BlockT::iterator I) {
+  while (I--) {
+    if (I->getOpcode() == AMDGPU::PRED_X) {
+      switch (static_cast<MachineInstr *>(I)->getOperand(2).getImm()) {
+      case OPCODE_IS_ZERO_INT:
+        static_cast<MachineInstr *>(I)->getOperand(2).setImm(OPCODE_IS_NOT_ZERO_INT);
+        return;
+      case OPCODE_IS_NOT_ZERO_INT:
+        static_cast<MachineInstr *>(I)->getOperand(2).setImm(OPCODE_IS_ZERO_INT);
+        return;
+      case OPCODE_IS_ZERO:
+        static_cast<MachineInstr *>(I)->getOperand(2).setImm(OPCODE_IS_NOT_ZERO);
+        return;
+      case OPCODE_IS_NOT_ZERO:
+        static_cast<MachineInstr *>(I)->getOperand(2).setImm(OPCODE_IS_ZERO);
+        return;
+      default:
+        assert(0 && "PRED_X Opcode invalid!");
+      }
+    }
+  }
+}
+
+template<class PassT>
+void CFGStructurizer<PassT>::mergeLoopbreakBlock(BlockT *exitingBlk,
+                                                 BlockT *exitBlk,
+                                                 BlockT *exitLandBlk,
+                                                 RegiT  setReg) {
+  if (DEBUGME) {
+    errs() << "loopbreakPattern exiting = BB" << exitingBlk->getNumber()
+           << " exit = BB" << exitBlk->getNumber()
+           << " land = BB" << exitLandBlk->getNumber() << "\n";
+  }
+
+  InstrT *branchInstr = CFGTraits::getLoopendBlockBranchInstr(exitingBlk);
+  assert(branchInstr && CFGTraits::isCondBranch(branchInstr));
+
+  DebugLoc DL = branchInstr->getDebugLoc();
+
+  BlockT *trueBranch = CFGTraits::getTrueBranch(branchInstr);
+
+  //    transform exitingBlk to
+  //    if ( ) {
+  //       exitBlk (if exitBlk != exitLandBlk)
+  //       setReg = 1
+  //       break
+  //    }endif
+  //    successor = {orgSuccessor(exitingBlk) - exitBlk}
+
+  typename BlockT::iterator branchInstrPos =
+    CFGTraits::getInstrPos(exitingBlk, branchInstr);
+
+  if (exitBlk == exitLandBlk && setReg == INVALIDREGNUM) {
+    //break_logical
+
+    if (trueBranch != exitBlk) {
+      reversePredicateSetter(branchInstrPos);
+    }
+    CFGTraits::insertCondBranchBefore(branchInstrPos, AMDGPU::PREDICATED_BREAK, passRep, DL);
+  } else {
+    if (trueBranch != exitBlk) {
+      reversePredicateSetter(branchInstr);
+    }
+    CFGTraits::insertCondBranchBefore(branchInstrPos, AMDGPU::PREDICATED_BREAK, passRep, DL);
+    if (exitBlk != exitLandBlk) {
+      //splice is insert-before ...
+      exitingBlk->splice(branchInstrPos, exitBlk, exitBlk->begin(),
+                         exitBlk->end());
+    }
+    if (setReg != INVALIDREGNUM) {
+      CFGTraits::insertAssignInstrBefore(branchInstrPos, passRep, setReg, 1);
+    }
+    CFGTraits::insertInstrBefore(branchInstrPos, AMDGPU::BREAK, passRep);
+  } //if_logical
+
+  //now branchInst can be erase safely
+  branchInstr->eraseFromParent();
+
+  //now take care of successors, retire blocks
+  exitingBlk->removeSuccessor(exitBlk);
+  if (exitBlk != exitLandBlk) {
+    //splice is insert-before ...
+    exitBlk->removeSuccessor(exitLandBlk);
+    retireBlock(exitingBlk, exitBlk);
+  }
+
+} //mergeLoopbreakBlock
+
+template<class PassT>
+void CFGStructurizer<PassT>::settleLoopcontBlock(BlockT *contingBlk,
+                                                 BlockT *contBlk,
+                                                 RegiT   setReg) {
+  if (DEBUGME) {
+    errs() << "settleLoopcontBlock conting = BB"
+           << contingBlk->getNumber()
+           << ", cont = BB" << contBlk->getNumber() << "\n";
+  }
+
+  InstrT *branchInstr = CFGTraits::getLoopendBlockBranchInstr(contingBlk);
+  if (branchInstr) {
+    assert(CFGTraits::isCondBranch(branchInstr));
+    typename BlockT::iterator branchInstrPos =
+      CFGTraits::getInstrPos(contingBlk, branchInstr);
+    BlockT *trueBranch = CFGTraits::getTrueBranch(branchInstr);
+    int oldOpcode = branchInstr->getOpcode();
+    DebugLoc DL = branchInstr->getDebugLoc();
+
+    //    transform contingBlk to
+    //     if () {
+    //          move instr after branchInstr
+    //          continue
+    //        or
+    //          setReg = 1
+    //          break
+    //     }endif
+    //     successor = {orgSuccessor(contingBlk) - loopHeader}
+
+    bool useContinueLogical = 
+      (setReg == INVALIDREGNUM && (&*contingBlk->rbegin()) == branchInstr);
+
+    if (useContinueLogical == false) {
+      int branchOpcode =
+        trueBranch == contBlk ? CFGTraits::getBranchNzeroOpcode(oldOpcode)
+                              : CFGTraits::getBranchZeroOpcode(oldOpcode);
+
+      CFGTraits::insertCondBranchBefore(branchInstrPos, branchOpcode, passRep, DL);
+
+      if (setReg != INVALIDREGNUM) {
+        CFGTraits::insertAssignInstrBefore(branchInstrPos, passRep, setReg, 1);
+        // insertEnd to ensure phi-moves, if exist, go before the continue-instr.
+        CFGTraits::insertInstrEnd(contingBlk, AMDGPU::BREAK, passRep, DL);
+      } else {
+        // insertEnd to ensure phi-moves, if exist, go before the continue-instr.
+        CFGTraits::insertInstrEnd(contingBlk, AMDGPU::CONTINUE, passRep, DL);
+      }
+
+      CFGTraits::insertInstrEnd(contingBlk, AMDGPU::ENDIF, passRep, DL);
+    } else {
+      int branchOpcode =
+        trueBranch == contBlk ? CFGTraits::getContinueNzeroOpcode(oldOpcode)
+                              : CFGTraits::getContinueZeroOpcode(oldOpcode);
+
+      CFGTraits::insertCondBranchBefore(branchInstrPos, branchOpcode, passRep, DL);
+    }
+
+    branchInstr->eraseFromParent();
+  } else {
+    // if we've arrived here then we've already erased the branch instruction
+    // travel back up the basic block to see the last reference of our debug location
+    // we've just inserted that reference here so it should be representative
+    if (setReg != INVALIDREGNUM) {
+      CFGTraits::insertAssignInstrBefore(contingBlk, passRep, setReg, 1);
+      // insertEnd to ensure phi-moves, if exist, go before the continue-instr.
+      CFGTraits::insertInstrEnd(contingBlk, AMDGPU::BREAK, passRep, CFGTraits::getLastDebugLocInBB(contingBlk));
+    } else {
+      // insertEnd to ensure phi-moves, if exist, go before the continue-instr.
+      CFGTraits::insertInstrEnd(contingBlk, AMDGPU::CONTINUE, passRep, CFGTraits::getLastDebugLocInBB(contingBlk));
+    }
+  } //else
+
+} //settleLoopcontBlock
+
+// BBs in exitBlkSet are determined as in break-path for loopRep,
+// before we can put code for BBs as inside loop-body for loopRep
+// check whether those BBs are determined as cont-BB for parentLoopRep
+// earlier.
+// If so, generate a new BB newBlk
+//    (1) set newBlk common successor of BBs in exitBlkSet
+//    (2) change the continue-instr in BBs in exitBlkSet to break-instr
+//    (3) generate continue-instr in newBlk
+//
+template<class PassT>
+typename CFGStructurizer<PassT>::BlockT *
+CFGStructurizer<PassT>::relocateLoopcontBlock(LoopT *parentLoopRep,
+                                              LoopT *loopRep,
+                                              std::set<BlockT *> &exitBlkSet,
+                                              BlockT *exitLandBlk) {
+  std::set<BlockT *> endBlkSet;
+
+
+
+  for (typename std::set<BlockT *>::const_iterator iter = exitBlkSet.begin(),
+       iterEnd = exitBlkSet.end();
+       iter != iterEnd; ++iter) {
+    BlockT *exitBlk = *iter;
+    BlockT *endBlk = singlePathEnd(exitBlk, exitLandBlk);
+
+    if (endBlk == NULL || CFGTraits::getContinueInstr(endBlk) == NULL)
+      return NULL;
+
+    endBlkSet.insert(endBlk);
+  }
+
+  BlockT *newBlk = funcRep->CreateMachineBasicBlock();
+  funcRep->push_back(newBlk);  //insert to function
+  CFGTraits::insertInstrEnd(newBlk, AMDGPU::CONTINUE, passRep);
+  SHOWNEWBLK(newBlk, "New continue block: ");
+
+  for (typename std::set<BlockT*>::const_iterator iter = endBlkSet.begin(),
+       iterEnd = endBlkSet.end();
+       iter != iterEnd; ++iter) {
+      BlockT *endBlk = *iter;
+      InstrT *contInstr = CFGTraits::getContinueInstr(endBlk);
+      if (contInstr) {
+        contInstr->eraseFromParent();
+      }
+      endBlk->addSuccessor(newBlk);
+      if (DEBUGME) {
+        errs() << "Add new continue Block to BB"
+               << endBlk->getNumber() << " successors\n";
+      }
+  }
+
+  return newBlk;
+} //relocateLoopcontBlock
+
+
+// LoopEndbranchBlock is a BB created by the CFGStructurizer to use as
+// LoopLandBlock. This BB branch on the loop endBranchInit register to the
+// pathes corresponding to the loop exiting branches.
+
+template<class PassT>
+typename CFGStructurizer<PassT>::BlockT *
+CFGStructurizer<PassT>::addLoopEndbranchBlock(LoopT *loopRep,
+                                              BlockTSmallerVector &exitingBlks,
+                                              BlockTSmallerVector &exitBlks) {
+  const AMDGPUInstrInfo *tii =
+             static_cast<const AMDGPUInstrInfo *>(passRep->getTargetInstrInfo());
+  const TargetRegisterClass * I32RC = TRI->getCFGStructurizerRegClass(MVT::i32);
+
+  RegiT endBranchReg = static_cast<int>
+    (funcRep->getRegInfo().createVirtualRegister(I32RC));
+  assert(endBranchReg >= 0);
+
+  // reg = 0 before entering the loop
+  addLoopEndbranchInitReg(loopRep, endBranchReg);
+
+  uint32_t numBlks = static_cast<uint32_t>(exitingBlks.size());
+  assert(numBlks >=2 && numBlks == exitBlks.size());
+
+  BlockT *preExitingBlk = exitingBlks[0];
+  BlockT *preExitBlk = exitBlks[0];
+  BlockT *preBranchBlk = funcRep->CreateMachineBasicBlock();
+  funcRep->push_back(preBranchBlk);  //insert to function
+  SHOWNEWBLK(preBranchBlk, "New loopEndbranch block: ");
+
+  BlockT *newLandBlk = preBranchBlk;
+
+      CFGTraits::replaceInstrUseOfBlockWith(preExitingBlk, preExitBlk,
+        newLandBlk);
+  preExitingBlk->removeSuccessor(preExitBlk);
+  preExitingBlk->addSuccessor(newLandBlk);
+
+  //it is redundant to add reg = 0 to exitingBlks[0]
+
+  // For 1..n th exiting path (the last iteration handles two pathes) create the
+  // branch to the previous path and the current path.
+  for (uint32_t i = 1; i < numBlks; ++i) {
+    BlockT *curExitingBlk = exitingBlks[i];
+    BlockT *curExitBlk = exitBlks[i];
+    BlockT *curBranchBlk;
+
+    if (i == numBlks - 1) {
+      curBranchBlk = curExitBlk;
+    } else {
+      curBranchBlk = funcRep->CreateMachineBasicBlock();
+      funcRep->push_back(curBranchBlk);  //insert to function
+      SHOWNEWBLK(curBranchBlk, "New loopEndbranch block: ");
+    }
+
+    // Add reg = i to exitingBlks[i].
+    CFGTraits::insertAssignInstrBefore(curExitingBlk, passRep,
+                                       endBranchReg, i);
+
+    // Remove the edge (exitingBlks[i] exitBlks[i]) add new edge
+    // (exitingBlks[i], newLandBlk).
+    CFGTraits::replaceInstrUseOfBlockWith(curExitingBlk, curExitBlk,
+                                          newLandBlk);
+    curExitingBlk->removeSuccessor(curExitBlk);
+    curExitingBlk->addSuccessor(newLandBlk);
+
+    // add to preBranchBlk the branch instruction:
+    // if (endBranchReg == preVal)
+    //    preExitBlk
+    // else
+    //    curBranchBlk
+    //
+    // preValReg = i - 1
+
+  DebugLoc DL;
+  RegiT preValReg = static_cast<int>
+    (funcRep->getRegInfo().createVirtualRegister(I32RC));
+
+  preBranchBlk->insert(preBranchBlk->begin(),
+                       tii->getMovImmInstr(preBranchBlk->getParent(), preValReg,
+                       i - 1));
+
+  // condResReg = (endBranchReg == preValReg)
+    RegiT condResReg = static_cast<int>
+      (funcRep->getRegInfo().createVirtualRegister(I32RC));
+    BuildMI(preBranchBlk, DL, tii->get(tii->getIEQOpcode()), condResReg)
+      .addReg(endBranchReg).addReg(preValReg);
+
+    BuildMI(preBranchBlk, DL, tii->get(AMDGPU::BRANCH_COND_i32))
+      .addMBB(preExitBlk).addReg(condResReg);
+
+    preBranchBlk->addSuccessor(preExitBlk);
+    preBranchBlk->addSuccessor(curBranchBlk);
+
+    // Update preExitingBlk, preExitBlk, preBranchBlk.
+    preExitingBlk = curExitingBlk;
+    preExitBlk = curExitBlk;
+    preBranchBlk = curBranchBlk;
+
+  }  //end for 1 .. n blocks
+
+  return newLandBlk;
+} //addLoopEndbranchBlock
+
+template<class PassT>
+typename CFGStructurizer<PassT>::PathToKind
+CFGStructurizer<PassT>::singlePathTo(BlockT *srcBlk, BlockT *dstBlk,
+                                     bool allowSideEntry) {
+  assert(dstBlk);
+
+  if (srcBlk == dstBlk) {
+    return SinglePath_InPath;
+  }
+
+  while (srcBlk && srcBlk->succ_size() == 1) {
+    srcBlk = *srcBlk->succ_begin();
+    if (srcBlk == dstBlk) {
+      return SinglePath_InPath;
+    }
+
+    if (!allowSideEntry && srcBlk->pred_size() > 1) {
+      return Not_SinglePath;
+    }
+  }
+
+  if (srcBlk && srcBlk->succ_size()==0) {
+    return SinglePath_NotInPath;
+  }
+
+  return Not_SinglePath;
+} //singlePathTo
+
+// If there is a single path from srcBlk to dstBlk, return the last block before
+// dstBlk If there is a single path from srcBlk->end without dstBlk, return the
+// last block in the path Otherwise, return NULL
+template<class PassT>
+typename CFGStructurizer<PassT>::BlockT *
+CFGStructurizer<PassT>::singlePathEnd(BlockT *srcBlk, BlockT *dstBlk,
+                                      bool allowSideEntry) {
+  assert(dstBlk);
+
+  if (srcBlk == dstBlk) {
+    return srcBlk;
+  }
+
+  if (srcBlk->succ_size() == 0) {
+    return srcBlk;
+  }
+
+  while (srcBlk && srcBlk->succ_size() == 1) {
+    BlockT *preBlk = srcBlk;
+
+    srcBlk = *srcBlk->succ_begin();
+    if (srcBlk == NULL) {
+      return preBlk;
+    }
+
+    if (!allowSideEntry && srcBlk->pred_size() > 1) {
+      return NULL;
+    }
+  }
+
+  if (srcBlk && srcBlk->succ_size()==0) {
+    return srcBlk;
+  }
+
+  return NULL;
+
+} //singlePathEnd
+
+template<class PassT>
+int CFGStructurizer<PassT>::cloneOnSideEntryTo(BlockT *preBlk, BlockT *srcBlk,
+                                               BlockT *dstBlk) {
+  int cloned = 0;
+  assert(preBlk->isSuccessor(srcBlk));
+  while (srcBlk && srcBlk != dstBlk) {
+    assert(srcBlk->succ_size() == 1);
+    if (srcBlk->pred_size() > 1) {
+      srcBlk = cloneBlockForPredecessor(srcBlk, preBlk);
+      ++cloned;
+    }
+
+    preBlk = srcBlk;
+    srcBlk = *srcBlk->succ_begin();
+  }
+
+  return cloned;
+} //cloneOnSideEntryTo
+
+template<class PassT>
+typename CFGStructurizer<PassT>::BlockT *
+CFGStructurizer<PassT>::cloneBlockForPredecessor(BlockT *curBlk,
+                                                 BlockT *predBlk) {
+  assert(predBlk->isSuccessor(curBlk) &&
+         "succBlk is not a prececessor of curBlk");
+
+  BlockT *cloneBlk = CFGTraits::clone(curBlk);  //clone instructions
+  CFGTraits::replaceInstrUseOfBlockWith(predBlk, curBlk, cloneBlk);
+  //srcBlk, oldBlk, newBlk
+
+  predBlk->removeSuccessor(curBlk);
+  predBlk->addSuccessor(cloneBlk);
+
+  // add all successor to cloneBlk
+  CFGTraits::cloneSuccessorList(cloneBlk, curBlk);
+
+  numClonedInstr += curBlk->size();
+
+  if (DEBUGME) {
+    errs() << "Cloned block: " << "BB"
+           << curBlk->getNumber() << "size " << curBlk->size() << "\n";
+  }
+
+  SHOWNEWBLK(cloneBlk, "result of Cloned block: ");
+
+  return cloneBlk;
+} //cloneBlockForPredecessor
+
+template<class PassT>
+typename CFGStructurizer<PassT>::BlockT *
+CFGStructurizer<PassT>::exitingBlock2ExitBlock(LoopT *loopRep,
+                                               BlockT *exitingBlk) {
+  BlockT *exitBlk = NULL;
+
+  for (typename BlockT::succ_iterator iterSucc = exitingBlk->succ_begin(),
+       iterSuccEnd = exitingBlk->succ_end();
+       iterSucc != iterSuccEnd; ++iterSucc) {
+    BlockT *curBlk = *iterSucc;
+    if (!loopRep->contains(curBlk)) {
+      assert(exitBlk == NULL);
+      exitBlk = curBlk;
+    }
+  }
+
+  assert(exitBlk != NULL);
+
+  return exitBlk;
+} //exitingBlock2ExitBlock
+
+template<class PassT>
+void CFGStructurizer<PassT>::migrateInstruction(BlockT *srcBlk,
+                                                BlockT *dstBlk,
+                                                InstrIterator insertPos) {
+  InstrIterator spliceEnd;
+  //look for the input branchinstr, not the AMDGPU branchinstr
+  InstrT *branchInstr = CFGTraits::getNormalBlockBranchInstr(srcBlk);
+  if (branchInstr == NULL) {
+    if (DEBUGME) {
+      errs() << "migrateInstruction don't see branch instr\n" ;
+    }
+    spliceEnd = srcBlk->end();
+  } else {
+    if (DEBUGME) {
+      errs() << "migrateInstruction see branch instr\n" ;
+      branchInstr->dump();
+    }
+    spliceEnd = CFGTraits::getInstrPos(srcBlk, branchInstr);
+  }
+  if (DEBUGME) {
+    errs() << "migrateInstruction before splice dstSize = " << dstBlk->size()
+      << "srcSize = " << srcBlk->size() << "\n";
+  }
+
+  //splice insert before insertPos
+  dstBlk->splice(insertPos, srcBlk, srcBlk->begin(), spliceEnd);
+
+  if (DEBUGME) {
+    errs() << "migrateInstruction after splice dstSize = " << dstBlk->size()
+      << "srcSize = " << srcBlk->size() << "\n";
+  }
+} //migrateInstruction
+
+// normalizeInfiniteLoopExit change
+//   B1:
+//        uncond_br LoopHeader
+//
+// to
+//   B1:
+//        cond_br 1 LoopHeader dummyExit
+// and return the newly added dummy exit block
+// 
+template<class PassT>
+typename CFGStructurizer<PassT>::BlockT *
+CFGStructurizer<PassT>::normalizeInfiniteLoopExit(LoopT* LoopRep) {
+  BlockT *loopHeader;
+  BlockT *loopLatch;
+  loopHeader = LoopRep->getHeader();
+  loopLatch = LoopRep->getLoopLatch();
+  BlockT *dummyExitBlk = NULL;
+  const TargetRegisterClass * I32RC = TRI->getCFGStructurizerRegClass(MVT::i32);
+  if (loopHeader!=NULL && loopLatch!=NULL) {
+    InstrT *branchInstr = CFGTraits::getLoopendBlockBranchInstr(loopLatch);
+    if (branchInstr!=NULL && CFGTraits::isUncondBranch(branchInstr)) {
+      dummyExitBlk = funcRep->CreateMachineBasicBlock();
+      funcRep->push_back(dummyExitBlk);  //insert to function
+      SHOWNEWBLK(dummyExitBlk, "DummyExitBlock to normalize infiniteLoop: ");
+
+      if (DEBUGME) errs() << "Old branch instr: " << *branchInstr << "\n";
+
+      typename BlockT::iterator insertPos =
+        CFGTraits::getInstrPos(loopLatch, branchInstr);
+      unsigned immReg =
+        funcRep->getRegInfo().createVirtualRegister(I32RC);
+      CFGTraits::insertAssignInstrBefore(insertPos, passRep, immReg, 1);
+      InstrT *newInstr = 
+        CFGTraits::insertInstrBefore(insertPos, AMDGPU::BRANCH_COND_i32, passRep);
+      MachineInstrBuilder(newInstr).addMBB(loopHeader).addReg(immReg, false);
+
+      SHOWNEWINSTR(newInstr);
+
+      branchInstr->eraseFromParent();
+      loopLatch->addSuccessor(dummyExitBlk);
+    }
+  }
+
+  return dummyExitBlk;
+} //normalizeInfiniteLoopExit
+
+template<class PassT>
+void CFGStructurizer<PassT>::removeUnconditionalBranch(BlockT *srcBlk) {
+  InstrT *branchInstr;
+
+  // I saw two unconditional branch in one basic block in example
+  // test_fc_do_while_or.c need to fix the upstream on this to remove the loop.
+  while ((branchInstr = CFGTraits::getLoopendBlockBranchInstr(srcBlk))
+          && CFGTraits::isUncondBranch(branchInstr)) {
+    if (DEBUGME) {
+          errs() << "Removing unconditional branch instruction" ;
+      branchInstr->dump();
+    }
+    branchInstr->eraseFromParent();
+  }
+} //removeUnconditionalBranch
+
+template<class PassT>
+void CFGStructurizer<PassT>::removeRedundantConditionalBranch(BlockT *srcBlk) {
+  if (srcBlk->succ_size() == 2) {
+    BlockT *blk1 = *srcBlk->succ_begin();
+    BlockT *blk2 = *(++srcBlk->succ_begin());
+
+    if (blk1 == blk2) {
+      InstrT *branchInstr = CFGTraits::getNormalBlockBranchInstr(srcBlk);
+      assert(branchInstr && CFGTraits::isCondBranch(branchInstr));
+      if (DEBUGME) {
+        errs() << "Removing unneeded conditional branch instruction" ;
+        branchInstr->dump();
+      }
+      branchInstr->eraseFromParent();
+      SHOWNEWBLK(blk1, "Removing redundant successor");
+      srcBlk->removeSuccessor(blk1);
+    }
+  }
+} //removeRedundantConditionalBranch
+
+template<class PassT>
+void CFGStructurizer<PassT>::addDummyExitBlock(SmallVector<BlockT*,
+                                               DEFAULT_VEC_SLOTS> &retBlks) {
+  BlockT *dummyExitBlk = funcRep->CreateMachineBasicBlock();
+  funcRep->push_back(dummyExitBlk);  //insert to function
+  CFGTraits::insertInstrEnd(dummyExitBlk, AMDGPU::RETURN, passRep);
+
+  for (typename SmallVector<BlockT *, DEFAULT_VEC_SLOTS>::iterator iter =
+         retBlks.begin(),
+       iterEnd = retBlks.end(); iter != iterEnd; ++iter) {
+    BlockT *curBlk = *iter;
+    InstrT *curInstr = CFGTraits::getReturnInstr(curBlk);
+    if (curInstr) {
+      curInstr->eraseFromParent();
+    }
+    curBlk->addSuccessor(dummyExitBlk);
+    if (DEBUGME) {
+      errs() << "Add dummyExitBlock to BB" << curBlk->getNumber()
+             << " successors\n";
+    }
+  } //for
+
+  SHOWNEWBLK(dummyExitBlk, "DummyExitBlock: ");
+} //addDummyExitBlock
+
+template<class PassT>
+void CFGStructurizer<PassT>::removeSuccessor(BlockT *srcBlk) {
+  while (srcBlk->succ_size()) {
+    srcBlk->removeSuccessor(*srcBlk->succ_begin());
+  }
+}
+
+template<class PassT>
+void CFGStructurizer<PassT>::recordSccnum(BlockT *srcBlk, int sccNum) {
+  BlockInfo *&srcBlkInfo = blockInfoMap[srcBlk];
+
+  if (srcBlkInfo == NULL) {
+    srcBlkInfo = new BlockInfo();
+  }
+
+  srcBlkInfo->sccNum = sccNum;
+}
+
+template<class PassT>
+int CFGStructurizer<PassT>::getSCCNum(BlockT *srcBlk) {
+  BlockInfo *srcBlkInfo = blockInfoMap[srcBlk];
+  return srcBlkInfo ? srcBlkInfo->sccNum : INVALIDSCCNUM;
+}
+
+template<class PassT>
+void CFGStructurizer<PassT>::retireBlock(BlockT *dstBlk, BlockT *srcBlk) {
+  if (DEBUGME) {
+        errs() << "Retiring BB" << srcBlk->getNumber() << "\n";
+  }
+
+  BlockInfo *&srcBlkInfo = blockInfoMap[srcBlk];
+
+  if (srcBlkInfo == NULL) {
+    srcBlkInfo = new BlockInfo();
+  }
+
+  srcBlkInfo->isRetired = true;
+  assert(srcBlk->succ_size() == 0 && srcBlk->pred_size() == 0
+         && "can't retire block yet");
+}
+
+template<class PassT>
+bool CFGStructurizer<PassT>::isRetiredBlock(BlockT *srcBlk) {
+  BlockInfo *srcBlkInfo = blockInfoMap[srcBlk];
+  return (srcBlkInfo && srcBlkInfo->isRetired);
+}
+
+template<class PassT>
+bool CFGStructurizer<PassT>::isActiveLoophead(BlockT *curBlk) {
+  LoopT *loopRep = loopInfo->getLoopFor(curBlk);
+  while (loopRep && loopRep->getHeader() == curBlk) {
+    LoopLandInfo *loopLand = getLoopLandInfo(loopRep);
+
+    if(loopLand == NULL)
+      return true;
+
+    BlockT *landBlk = loopLand->landBlk;
+    assert(landBlk);
+    if (!isRetiredBlock(landBlk)) {
+      return true;
+    }
+
+    loopRep = loopRep->getParentLoop();
+  }
+
+  return false;
+} //isActiveLoophead
+
+template<class PassT>
+bool CFGStructurizer<PassT>::needMigrateBlock(BlockT *blk) {
+  const unsigned blockSizeThreshold = 30;
+  const unsigned cloneInstrThreshold = 100;
+
+  bool multiplePreds = blk && (blk->pred_size() > 1);
+
+  if(!multiplePreds)
+    return false;
+
+  unsigned blkSize = blk->size();
+  return ((blkSize > blockSizeThreshold)
+          && (blkSize * (blk->pred_size() - 1) > cloneInstrThreshold));
+} //needMigrateBlock
+
+template<class PassT>
+typename CFGStructurizer<PassT>::BlockT *
+CFGStructurizer<PassT>::recordLoopLandBlock(LoopT *loopRep, BlockT *landBlk,
+                                            BlockTSmallerVector &exitBlks,
+                                            std::set<BlockT *> &exitBlkSet) {
+  SmallVector<BlockT *, DEFAULT_VEC_SLOTS> inpathBlks;  //in exit path blocks
+
+  for (typename BlockT::pred_iterator predIter = landBlk->pred_begin(),
+       predIterEnd = landBlk->pred_end();
+       predIter != predIterEnd; ++predIter) {
+    BlockT *curBlk = *predIter;
+    if (loopRep->contains(curBlk) || exitBlkSet.count(curBlk)) {
+      inpathBlks.push_back(curBlk);
+    }
+  } //for
+
+  //if landBlk has predecessors that are not in the given loop,
+  //create a new block
+  BlockT *newLandBlk = landBlk;
+  if (inpathBlks.size() != landBlk->pred_size()) {
+    newLandBlk = funcRep->CreateMachineBasicBlock();
+    funcRep->push_back(newLandBlk);  //insert to function
+    newLandBlk->addSuccessor(landBlk);
+    for (typename SmallVector<BlockT*, DEFAULT_VEC_SLOTS>::iterator iter =
+         inpathBlks.begin(),
+         iterEnd = inpathBlks.end(); iter != iterEnd; ++iter) {
+      BlockT *curBlk = *iter;
+      CFGTraits::replaceInstrUseOfBlockWith(curBlk, landBlk, newLandBlk);
+      //srcBlk, oldBlk, newBlk
+      curBlk->removeSuccessor(landBlk);
+      curBlk->addSuccessor(newLandBlk);
+    }
+    for (size_t i = 0, tot = exitBlks.size(); i < tot; ++i) {
+      if (exitBlks[i] == landBlk) {
+        exitBlks[i] = newLandBlk;
+      }
+    }
+    SHOWNEWBLK(newLandBlk, "NewLandingBlock: ");
+  }
+
+  setLoopLandBlock(loopRep, newLandBlk);
+
+  return newLandBlk;
+} // recordLoopbreakLand
+
+template<class PassT>
+void CFGStructurizer<PassT>::setLoopLandBlock(LoopT *loopRep, BlockT *blk) {
+  LoopLandInfo *&theEntry = loopLandInfoMap[loopRep];
+
+  if (theEntry == NULL) {
+    theEntry = new LoopLandInfo();
+  }
+  assert(theEntry->landBlk == NULL);
+
+  if (blk == NULL) {
+    blk = funcRep->CreateMachineBasicBlock();
+    funcRep->push_back(blk);  //insert to function
+    SHOWNEWBLK(blk, "DummyLandingBlock for loop without break: ");
+  }
+
+  theEntry->landBlk = blk;
+
+  if (DEBUGME) {
+    errs() << "setLoopLandBlock loop-header = BB"
+           << loopRep->getHeader()->getNumber()
+           << "  landing-block = BB" << blk->getNumber() << "\n";
+  }
+} // setLoopLandBlock
+
+template<class PassT>
+void CFGStructurizer<PassT>::addLoopBreakOnReg(LoopT *loopRep, RegiT regNum) {
+  LoopLandInfo *&theEntry = loopLandInfoMap[loopRep];
+
+  if (theEntry == NULL) {
+    theEntry = new LoopLandInfo();
+  }
+
+  theEntry->breakOnRegs.insert(regNum);
+
+  if (DEBUGME) {
+    errs() << "addLoopBreakOnReg loop-header = BB"
+           << loopRep->getHeader()->getNumber()
+           << "  regNum = " << regNum << "\n";
+  }
+} // addLoopBreakOnReg
+
+template<class PassT>
+void CFGStructurizer<PassT>::addLoopContOnReg(LoopT *loopRep, RegiT regNum) {
+  LoopLandInfo *&theEntry = loopLandInfoMap[loopRep];
+
+  if (theEntry == NULL) {
+    theEntry = new LoopLandInfo();
+  }
+  theEntry->contOnRegs.insert(regNum);
+
+  if (DEBUGME) {
+    errs() << "addLoopContOnReg loop-header = BB"
+           << loopRep->getHeader()->getNumber()
+           << "  regNum = " << regNum << "\n";
+  }
+} // addLoopContOnReg
+
+template<class PassT>
+void CFGStructurizer<PassT>::addLoopBreakInitReg(LoopT *loopRep, RegiT regNum) {
+  LoopLandInfo *&theEntry = loopLandInfoMap[loopRep];
+
+  if (theEntry == NULL) {
+    theEntry = new LoopLandInfo();
+  }
+  theEntry->breakInitRegs.insert(regNum);
+
+  if (DEBUGME) {
+    errs() << "addLoopBreakInitReg loop-header = BB"
+           << loopRep->getHeader()->getNumber()
+           << "  regNum = " << regNum << "\n";
+  }
+} // addLoopBreakInitReg
+
+template<class PassT>
+void CFGStructurizer<PassT>::addLoopContInitReg(LoopT *loopRep, RegiT regNum) {
+  LoopLandInfo *&theEntry = loopLandInfoMap[loopRep];
+
+  if (theEntry == NULL) {
+    theEntry = new LoopLandInfo();
+  }
+  theEntry->contInitRegs.insert(regNum);
+
+  if (DEBUGME) {
+    errs() << "addLoopContInitReg loop-header = BB"
+           << loopRep->getHeader()->getNumber()
+           << "  regNum = " << regNum << "\n";
+  }
+} // addLoopContInitReg
+
+template<class PassT>
+void CFGStructurizer<PassT>::addLoopEndbranchInitReg(LoopT *loopRep,
+                                                     RegiT regNum) {
+  LoopLandInfo *&theEntry = loopLandInfoMap[loopRep];
+
+  if (theEntry == NULL) {
+    theEntry = new LoopLandInfo();
+  }
+  theEntry->endbranchInitRegs.insert(regNum);
+
+  if (DEBUGME) {
+        errs() << "addLoopEndbranchInitReg loop-header = BB"
+      << loopRep->getHeader()->getNumber()
+      << "  regNum = " << regNum << "\n";
+  }
+} // addLoopEndbranchInitReg
+
+template<class PassT>
+typename CFGStructurizer<PassT>::LoopLandInfo *
+CFGStructurizer<PassT>::getLoopLandInfo(LoopT *loopRep) {
+  LoopLandInfo *&theEntry = loopLandInfoMap[loopRep];
+
+  return theEntry;
+} // getLoopLandInfo
+
+template<class PassT>
+typename CFGStructurizer<PassT>::BlockT *
+CFGStructurizer<PassT>::getLoopLandBlock(LoopT *loopRep) {
+  LoopLandInfo *&theEntry = loopLandInfoMap[loopRep];
+
+  return theEntry ? theEntry->landBlk : NULL;
+} // getLoopLandBlock
+
+
+template<class PassT>
+bool CFGStructurizer<PassT>::hasBackEdge(BlockT *curBlk) {
+  LoopT *loopRep = loopInfo->getLoopFor(curBlk);
+  if (loopRep == NULL)
+    return false;
+
+  BlockT *loopHeader = loopRep->getHeader();
+
+  return curBlk->isSuccessor(loopHeader);
+
+} //hasBackEdge
+
+template<class PassT>
+unsigned CFGStructurizer<PassT>::getLoopDepth(LoopT *loopRep) {
+  return loopRep ? loopRep->getLoopDepth() : 0;
+} //getLoopDepth
+
+template<class PassT>
+int CFGStructurizer<PassT>::countActiveBlock
+(typename SmallVector<BlockT*, DEFAULT_VEC_SLOTS>::const_iterator iterStart,
+ typename SmallVector<BlockT*, DEFAULT_VEC_SLOTS>::const_iterator iterEnd) {
+  int count = 0;
+  while (iterStart != iterEnd) {
+    if (!isRetiredBlock(*iterStart)) {
+      ++count;
+    }
+    ++iterStart;
+  }
+
+  return count;
+} //countActiveBlock
+
+// This is work around solution for findNearestCommonDominator not avaiable to
+// post dom a proper fix should go to Dominators.h.
+
+template<class PassT>
+typename CFGStructurizer<PassT>::BlockT*
+CFGStructurizer<PassT>::findNearestCommonPostDom(BlockT *blk1, BlockT *blk2) {
+
+  if (postDomTree->dominates(blk1, blk2)) {
+    return blk1;
+  }
+  if (postDomTree->dominates(blk2, blk1)) {
+    return blk2;
+  }
+
+  DomTreeNodeT *node1 = postDomTree->getNode(blk1);
+  DomTreeNodeT *node2 = postDomTree->getNode(blk2);
+
+  // Handle newly cloned node.
+  if (node1 == NULL && blk1->succ_size() == 1) {
+    return findNearestCommonPostDom(*blk1->succ_begin(), blk2);
+  }
+  if (node2 == NULL && blk2->succ_size() == 1) {
+    return findNearestCommonPostDom(blk1, *blk2->succ_begin());
+  }
+
+  if (node1 == NULL || node2 == NULL) {
+    return NULL;
+  }
+
+  node1 = node1->getIDom();
+  while (node1) {
+    if (postDomTree->dominates(node1, node2)) {
+      return node1->getBlock();
+    }
+    node1 = node1->getIDom();
+  }
+
+  return NULL;
+}
+
+template<class PassT>
+typename CFGStructurizer<PassT>::BlockT *
+CFGStructurizer<PassT>::findNearestCommonPostDom
+(typename std::set<BlockT *> &blks) {
+  BlockT *commonDom;
+  typename std::set<BlockT *>::const_iterator iter = blks.begin();
+  typename std::set<BlockT *>::const_iterator iterEnd = blks.end();
+  for (commonDom = *iter; iter != iterEnd && commonDom != NULL; ++iter) {
+    BlockT *curBlk = *iter;
+    if (curBlk != commonDom) {
+      commonDom = findNearestCommonPostDom(curBlk, commonDom);
+    }
+  }
+
+  if (DEBUGME) {
+    errs() << "Common post dominator for exit blocks is ";
+    if (commonDom) {
+          errs() << "BB" << commonDom->getNumber() << "\n";
+    } else {
+      errs() << "NULL\n";
+    }
+  }
+
+  return commonDom;
+} //findNearestCommonPostDom
+
+} //end namespace llvm
+
+//todo: move-end
+
+
+//===----------------------------------------------------------------------===//
+//
+// CFGStructurizer for AMDGPU
+//
+//===----------------------------------------------------------------------===//
+
+
+using namespace llvmCFGStruct;
+
+namespace llvm {
+class AMDGPUCFGStructurizer : public MachineFunctionPass {
+public:
+  typedef MachineInstr              InstructionType;
+  typedef MachineFunction           FunctionType;
+  typedef MachineBasicBlock         BlockType;
+  typedef MachineLoopInfo           LoopinfoType;
+  typedef MachineDominatorTree      DominatortreeType;
+  typedef MachinePostDominatorTree  PostDominatortreeType;
+  typedef MachineDomTreeNode        DomTreeNodeType;
+  typedef MachineLoop               LoopType;
+
+protected:
+  TargetMachine &TM;
+  const TargetInstrInfo *TII;
+  const AMDGPURegisterInfo *TRI;
+
+public:
+  AMDGPUCFGStructurizer(char &pid, TargetMachine &tm);
+  const TargetInstrInfo *getTargetInstrInfo() const;
+
+private:
+
+};
+
+} //end of namespace llvm
+AMDGPUCFGStructurizer::AMDGPUCFGStructurizer(char &pid, TargetMachine &tm)
+: MachineFunctionPass(pid), TM(tm), TII(tm.getInstrInfo()),
+  TRI(static_cast<const AMDGPURegisterInfo *>(tm.getRegisterInfo())) {
+}
+
+const TargetInstrInfo *AMDGPUCFGStructurizer::getTargetInstrInfo() const {
+  return TII;
+}
+//===----------------------------------------------------------------------===//
+//
+// CFGPrepare
+//
+//===----------------------------------------------------------------------===//
+
+
+using namespace llvmCFGStruct;
+
+namespace llvm {
+class AMDGPUCFGPrepare : public AMDGPUCFGStructurizer {
+public:
+  static char ID;
+
+public:
+  AMDGPUCFGPrepare(TargetMachine &tm);
+
+  virtual const char *getPassName() const;
+  virtual void getAnalysisUsage(AnalysisUsage &AU) const;
+
+  bool runOnMachineFunction(MachineFunction &F);
+
+private:
+
+};
+
+char AMDGPUCFGPrepare::ID = 0;
+} //end of namespace llvm
+
+AMDGPUCFGPrepare::AMDGPUCFGPrepare(TargetMachine &tm)
+  : AMDGPUCFGStructurizer(ID, tm )  {
+}
+const char *AMDGPUCFGPrepare::getPassName() const {
+  return "AMD IL Control Flow Graph Preparation Pass";
+}
+
+void AMDGPUCFGPrepare::getAnalysisUsage(AnalysisUsage &AU) const {
+  AU.addPreserved<MachineFunctionAnalysis>();
+  AU.addRequired<MachineFunctionAnalysis>();
+  AU.addRequired<MachineDominatorTree>();
+  AU.addRequired<MachinePostDominatorTree>();
+  AU.addRequired<MachineLoopInfo>();
+}
+
+//===----------------------------------------------------------------------===//
+//
+// CFGPerform
+//
+//===----------------------------------------------------------------------===//
+
+
+using namespace llvmCFGStruct;
+
+namespace llvm {
+class AMDGPUCFGPerform : public AMDGPUCFGStructurizer {
+public:
+  static char ID;
+
+public:
+  AMDGPUCFGPerform(TargetMachine &tm);
+  virtual const char *getPassName() const;
+  virtual void getAnalysisUsage(AnalysisUsage &AU) const;
+  bool runOnMachineFunction(MachineFunction &F);
+
+private:
+
+};
+
+char AMDGPUCFGPerform::ID = 0;
+} //end of namespace llvm
+
+  AMDGPUCFGPerform::AMDGPUCFGPerform(TargetMachine &tm)
+: AMDGPUCFGStructurizer(ID, tm) {
+}
+
+const char *AMDGPUCFGPerform::getPassName() const {
+  return "AMD IL Control Flow Graph structurizer Pass";
+}
+
+void AMDGPUCFGPerform::getAnalysisUsage(AnalysisUsage &AU) const {
+  AU.addPreserved<MachineFunctionAnalysis>();
+  AU.addRequired<MachineFunctionAnalysis>();
+  AU.addRequired<MachineDominatorTree>();
+  AU.addRequired<MachinePostDominatorTree>();
+  AU.addRequired<MachineLoopInfo>();
+}
+
+//===----------------------------------------------------------------------===//
+//
+// CFGStructTraits<AMDGPUCFGStructurizer>
+//
+//===----------------------------------------------------------------------===//
+
+namespace llvmCFGStruct {
+// this class is tailor to the AMDGPU backend
+template<>
+struct CFGStructTraits<AMDGPUCFGStructurizer> {
+  typedef int RegiT;
+
+  static int getBranchNzeroOpcode(int oldOpcode) {
+    switch(oldOpcode) {
+    case AMDGPU::JUMP: return AMDGPU::IF_PREDICATE_SET;
+    case AMDGPU::BRANCH_COND_i32:
+    case AMDGPU::BRANCH_COND_f32: return AMDGPU::IF_LOGICALNZ_f32;
+    case AMDGPU::SI_IF_NZ: return AMDGPU::SI_IF_NZ;
+    default:
+      assert(0 && "internal error");
+    }
+    return -1;
+  }
+
+  static int getBranchZeroOpcode(int oldOpcode) {
+    switch(oldOpcode) {
+    case AMDGPU::JUMP: return AMDGPU::IF_PREDICATE_SET;
+    case AMDGPU::BRANCH_COND_i32:
+    case AMDGPU::BRANCH_COND_f32: return AMDGPU::IF_LOGICALZ_f32;
+    case AMDGPU::SI_IF_Z: return AMDGPU::SI_IF_Z;
+    default:
+      assert(0 && "internal error");
+    }
+    return -1;
+  }
+
+  static int getContinueNzeroOpcode(int oldOpcode) {
+    switch(oldOpcode) {
+    case AMDGPU::JUMP: return AMDGPU::CONTINUE_LOGICALNZ_i32;
+    default:
+      assert(0 && "internal error");
+    };
+    return -1;
+  }
+
+  static int getContinueZeroOpcode(int oldOpcode) {
+    switch(oldOpcode) {
+    case AMDGPU::JUMP: return AMDGPU::CONTINUE_LOGICALZ_i32;
+    default:
+      assert(0 && "internal error");
+    }
+    return -1;
+  }
+
+  static MachineBasicBlock *getTrueBranch(MachineInstr *instr) {
+    return instr->getOperand(0).getMBB();
+  }
+
+  static void setTrueBranch(MachineInstr *instr, MachineBasicBlock *blk) {
+    instr->getOperand(0).setMBB(blk);
+  }
+
+  static MachineBasicBlock *
+  getFalseBranch(MachineBasicBlock *blk, MachineInstr *instr) {
+    assert(blk->succ_size() == 2);
+    MachineBasicBlock *trueBranch = getTrueBranch(instr);
+    MachineBasicBlock::succ_iterator iter = blk->succ_begin();
+    MachineBasicBlock::succ_iterator iterNext = iter;
+    ++iterNext;
+
+    return (*iter == trueBranch) ? *iterNext : *iter;
+  }
+
+  static bool isCondBranch(MachineInstr *instr) {
+    switch (instr->getOpcode()) {
+      case AMDGPU::JUMP:
+        return instr->getOperand(instr->findFirstPredOperandIdx()).getReg() != 0;
+      case AMDGPU::BRANCH_COND_i32:
+      case AMDGPU::BRANCH_COND_f32:
+      case AMDGPU::SI_IF_NZ:
+      case AMDGPU::SI_IF_Z:
+      break;
+    default:
+      return false;
+    }
+    return true;
+  }
+
+  static bool isUncondBranch(MachineInstr *instr) {
+    switch (instr->getOpcode()) {
+    case AMDGPU::JUMP:
+      return instr->getOperand(instr->findFirstPredOperandIdx()).getReg() == 0;
+    case AMDGPU::BRANCH:
+      return true;
+    default:
+      return false;
+    }
+    return true;
+  }
+
+  static DebugLoc getLastDebugLocInBB(MachineBasicBlock *blk) {
+    //get DebugLoc from the first MachineBasicBlock instruction with debug info
+    DebugLoc DL;
+    for (MachineBasicBlock::iterator iter = blk->begin(); iter != blk->end(); ++iter) {
+      MachineInstr *instr = &(*iter);
+      if (instr->getDebugLoc().isUnknown() == false) {
+        DL = instr->getDebugLoc();
+      }
+    }
+    return DL;
+  }
+
+  static MachineInstr *getNormalBlockBranchInstr(MachineBasicBlock *blk) {
+    MachineBasicBlock::reverse_iterator iter = blk->rbegin();
+    MachineInstr *instr = &*iter;
+    if (instr && (isCondBranch(instr) || isUncondBranch(instr))) {
+      return instr;
+    }
+    return NULL;
+  }
+
+  // The correct naming for this is getPossibleLoopendBlockBranchInstr.
+  //
+  // BB with backward-edge could have move instructions after the branch
+  // instruction.  Such move instruction "belong to" the loop backward-edge.
+  //
+  static MachineInstr *getLoopendBlockBranchInstr(MachineBasicBlock *blk) {
+    const AMDGPUInstrInfo * TII = static_cast<const AMDGPUInstrInfo *>(
+                                  blk->getParent()->getTarget().getInstrInfo());
+
+    for (MachineBasicBlock::reverse_iterator iter = blk->rbegin(),
+         iterEnd = blk->rend(); iter != iterEnd; ++iter) {
+      // FIXME: Simplify
+      MachineInstr *instr = &*iter;
+      if (instr) {
+        if (isCondBranch(instr) || isUncondBranch(instr)) {
+          return instr;
+        } else if (!TII->isMov(instr->getOpcode())) {
+          break;
+        }
+      }
+    }
+    return NULL;
+  }
+
+  static MachineInstr *getReturnInstr(MachineBasicBlock *blk) {
+    MachineBasicBlock::reverse_iterator iter = blk->rbegin();
+    if (iter != blk->rend()) {
+      MachineInstr *instr = &(*iter);
+      if (instr->getOpcode() == AMDGPU::RETURN) {
+        return instr;
+      }
+    }
+    return NULL;
+  }
+
+  static MachineInstr *getContinueInstr(MachineBasicBlock *blk) {
+    MachineBasicBlock::reverse_iterator iter = blk->rbegin();
+    if (iter != blk->rend()) {
+      MachineInstr *instr = &(*iter);
+      if (instr->getOpcode() == AMDGPU::CONTINUE) {
+        return instr;
+      }
+    }
+    return NULL;
+  }
+
+  static MachineInstr *getLoopBreakInstr(MachineBasicBlock *blk) {
+    for (MachineBasicBlock::iterator iter = blk->begin(); (iter != blk->end()); ++iter) {
+      MachineInstr *instr = &(*iter);
+      if (instr->getOpcode() == AMDGPU::PREDICATED_BREAK) {
+        return instr;
+      }
+    }
+    return NULL;
+  }
+
+  static bool isReturnBlock(MachineBasicBlock *blk) {
+    MachineInstr *instr = getReturnInstr(blk);
+    bool isReturn = (blk->succ_size() == 0);
+    if (instr) {
+      assert(isReturn);
+    } else if (isReturn) {
+      if (DEBUGME) {
+        errs() << "BB" << blk->getNumber()
+               <<" is return block without RETURN instr\n";
+      }
+    }
+
+    return  isReturn;
+  }
+
+  static MachineBasicBlock::iterator
+  getInstrPos(MachineBasicBlock *blk, MachineInstr *instr) {
+    assert(instr->getParent() == blk && "instruction doesn't belong to block");
+    MachineBasicBlock::iterator iter = blk->begin();
+    MachineBasicBlock::iterator iterEnd = blk->end();
+    while (&(*iter) != instr && iter != iterEnd) {
+      ++iter;
+    }
+
+    assert(iter != iterEnd);
+    return iter;
+  }//getInstrPos
+
+  static MachineInstr *insertInstrBefore(MachineBasicBlock *blk, int newOpcode,
+                                         AMDGPUCFGStructurizer *passRep) {
+    return insertInstrBefore(blk,newOpcode,passRep,DebugLoc());
+  } //insertInstrBefore
+
+  static MachineInstr *insertInstrBefore(MachineBasicBlock *blk, int newOpcode,
+                                         AMDGPUCFGStructurizer *passRep, DebugLoc DL) {
+    const TargetInstrInfo *tii = passRep->getTargetInstrInfo();
+    MachineInstr *newInstr =
+      blk->getParent()->CreateMachineInstr(tii->get(newOpcode), DL);
+
+    MachineBasicBlock::iterator res;
+    if (blk->begin() != blk->end()) {
+      blk->insert(blk->begin(), newInstr);
+    } else {
+      blk->push_back(newInstr);
+    }
+
+    SHOWNEWINSTR(newInstr);
+
+    return newInstr;
+  } //insertInstrBefore
+
+  static void insertInstrEnd(MachineBasicBlock *blk, int newOpcode,
+                             AMDGPUCFGStructurizer *passRep) {
+    insertInstrEnd(blk,newOpcode,passRep,DebugLoc());
+  } //insertInstrEnd
+
+  static void insertInstrEnd(MachineBasicBlock *blk, int newOpcode,
+                             AMDGPUCFGStructurizer *passRep, DebugLoc DL) {
+    const TargetInstrInfo *tii = passRep->getTargetInstrInfo();
+   MachineInstr *newInstr = blk->getParent()
+      ->CreateMachineInstr(tii->get(newOpcode), DL);
+
+    blk->push_back(newInstr);
+    //assume the instruction doesn't take any reg operand ...
+
+    SHOWNEWINSTR(newInstr);
+  } //insertInstrEnd
+
+  static MachineInstr *insertInstrBefore(MachineBasicBlock::iterator instrPos,
+                                         int newOpcode, 
+                                         AMDGPUCFGStructurizer *passRep) {
+    MachineInstr *oldInstr = &(*instrPos);
+    const TargetInstrInfo *tii = passRep->getTargetInstrInfo();
+    MachineBasicBlock *blk = oldInstr->getParent();
+    MachineInstr *newInstr =
+      blk->getParent()->CreateMachineInstr(tii->get(newOpcode),
+                                           DebugLoc());
+
+    blk->insert(instrPos, newInstr);
+    //assume the instruction doesn't take any reg operand ...
+
+    SHOWNEWINSTR(newInstr);
+    return newInstr;
+  } //insertInstrBefore
+
+  static void insertCondBranchBefore(MachineBasicBlock::iterator instrPos,
+                                     int newOpcode,
+                                     AMDGPUCFGStructurizer *passRep,
+                                     DebugLoc DL) {
+    MachineInstr *oldInstr = &(*instrPos);
+    const TargetInstrInfo *tii = passRep->getTargetInstrInfo();
+    MachineBasicBlock *blk = oldInstr->getParent();
+    MachineInstr *newInstr =
+      blk->getParent()->CreateMachineInstr(tii->get(newOpcode),
+                                           DL);
+
+    blk->insert(instrPos, newInstr);
+    MachineInstrBuilder(newInstr).addReg(oldInstr->getOperand(1).getReg(),
+                                         false);
+
+    SHOWNEWINSTR(newInstr);
+    //erase later oldInstr->eraseFromParent();
+  } //insertCondBranchBefore
+
+  static void insertCondBranchBefore(MachineBasicBlock *blk,
+                                     MachineBasicBlock::iterator insertPos,
+                                     int newOpcode,
+                                     AMDGPUCFGStructurizer *passRep,
+                                     RegiT regNum,
+                                     DebugLoc DL) {
+    const TargetInstrInfo *tii = passRep->getTargetInstrInfo();
+
+    MachineInstr *newInstr =
+      blk->getParent()->CreateMachineInstr(tii->get(newOpcode), DL);
+
+    //insert before
+    blk->insert(insertPos, newInstr);
+    MachineInstrBuilder(newInstr).addReg(regNum, false);
+
+    SHOWNEWINSTR(newInstr);
+  } //insertCondBranchBefore
+
+  static void insertCondBranchEnd(MachineBasicBlock *blk,
+                                  int newOpcode,
+                                  AMDGPUCFGStructurizer *passRep,
+                                  RegiT regNum) {
+    const TargetInstrInfo *tii = passRep->getTargetInstrInfo();
+    MachineInstr *newInstr =
+      blk->getParent()->CreateMachineInstr(tii->get(newOpcode), DebugLoc());
+
+    blk->push_back(newInstr);
+    MachineInstrBuilder(newInstr).addReg(regNum, false);
+
+    SHOWNEWINSTR(newInstr);
+  } //insertCondBranchEnd
+
+
+  static void insertAssignInstrBefore(MachineBasicBlock::iterator instrPos,
+                                      AMDGPUCFGStructurizer *passRep,
+                                      RegiT regNum, int regVal) {
+    MachineInstr *oldInstr = &(*instrPos);
+    const AMDGPUInstrInfo *tii =
+             static_cast<const AMDGPUInstrInfo *>(passRep->getTargetInstrInfo());
+    MachineBasicBlock *blk = oldInstr->getParent();
+    MachineInstr *newInstr = tii->getMovImmInstr(blk->getParent(), regNum,
+                                                 regVal);
+    blk->insert(instrPos, newInstr);
+
+    SHOWNEWINSTR(newInstr);
+  } //insertAssignInstrBefore
+
+  static void insertAssignInstrBefore(MachineBasicBlock *blk,
+                                      AMDGPUCFGStructurizer *passRep,
+                                      RegiT regNum, int regVal) {
+    const AMDGPUInstrInfo *tii =
+             static_cast<const AMDGPUInstrInfo *>(passRep->getTargetInstrInfo());
+
+    MachineInstr *newInstr = tii->getMovImmInstr(blk->getParent(), regNum,
+                                                 regVal);
+    if (blk->begin() != blk->end()) {
+      blk->insert(blk->begin(), newInstr);
+    } else {
+      blk->push_back(newInstr);
+    }
+
+    SHOWNEWINSTR(newInstr);
+
+  } //insertInstrBefore
+
+  static void insertCompareInstrBefore(MachineBasicBlock *blk,
+                                       MachineBasicBlock::iterator instrPos,
+                                       AMDGPUCFGStructurizer *passRep,
+                                       RegiT dstReg, RegiT src1Reg,
+                                       RegiT src2Reg) {
+    const AMDGPUInstrInfo *tii =
+             static_cast<const AMDGPUInstrInfo *>(passRep->getTargetInstrInfo());
+    MachineInstr *newInstr =
+      blk->getParent()->CreateMachineInstr(tii->get(tii->getIEQOpcode()), DebugLoc());
+
+    MachineInstrBuilder(newInstr).addReg(dstReg, RegState::Define); //set target
+    MachineInstrBuilder(newInstr).addReg(src1Reg); //set src value
+    MachineInstrBuilder(newInstr).addReg(src2Reg); //set src value
+
+    blk->insert(instrPos, newInstr);
+    SHOWNEWINSTR(newInstr);
+
+  } //insertCompareInstrBefore
+
+  static void cloneSuccessorList(MachineBasicBlock *dstBlk,
+                                 MachineBasicBlock *srcBlk) {
+    for (MachineBasicBlock::succ_iterator iter = srcBlk->succ_begin(),
+         iterEnd = srcBlk->succ_end(); iter != iterEnd; ++iter) {
+      dstBlk->addSuccessor(*iter);  // *iter's predecessor is also taken care of
+    }
+  } //cloneSuccessorList
+
+  static MachineBasicBlock *clone(MachineBasicBlock *srcBlk) {
+    MachineFunction *func = srcBlk->getParent();
+    MachineBasicBlock *newBlk = func->CreateMachineBasicBlock();
+    func->push_back(newBlk);  //insert to function
+    for (MachineBasicBlock::iterator iter = srcBlk->begin(),
+         iterEnd = srcBlk->end();
+         iter != iterEnd; ++iter) {
+      MachineInstr *instr = func->CloneMachineInstr(iter);
+      newBlk->push_back(instr);
+    }
+    return newBlk;
+  }
+
+  //MachineBasicBlock::ReplaceUsesOfBlockWith doesn't serve the purpose because
+  //the AMDGPU instruction is not recognized as terminator fix this and retire
+  //this routine
+  static void replaceInstrUseOfBlockWith(MachineBasicBlock *srcBlk,
+                                         MachineBasicBlock *oldBlk,
+                                         MachineBasicBlock *newBlk) {
+    MachineInstr *branchInstr = getLoopendBlockBranchInstr(srcBlk);
+    if (branchInstr && isCondBranch(branchInstr) &&
+        getTrueBranch(branchInstr) == oldBlk) {
+      setTrueBranch(branchInstr, newBlk);
+    }
+  }
+
+  static void wrapup(MachineBasicBlock *entryBlk) {
+    assert((!entryBlk->getParent()->getJumpTableInfo()
+            || entryBlk->getParent()->getJumpTableInfo()->isEmpty())
+           && "found a jump table");
+
+     //collect continue right before endloop
+     SmallVector<MachineInstr *, DEFAULT_VEC_SLOTS> contInstr;
+     MachineBasicBlock::iterator pre = entryBlk->begin();
+     MachineBasicBlock::iterator iterEnd = entryBlk->end();
+     MachineBasicBlock::iterator iter = pre;
+     while (iter != iterEnd) {
+       if (pre->getOpcode() == AMDGPU::CONTINUE
+           && iter->getOpcode() == AMDGPU::ENDLOOP) {
+         contInstr.push_back(pre);
+       }
+       pre = iter;
+       ++iter;
+     } //end while
+
+     //delete continue right before endloop
+     for (unsigned i = 0; i < contInstr.size(); ++i) {
+        contInstr[i]->eraseFromParent();
+     }
+
+     // TODO to fix up jump table so later phase won't be confused.  if
+     // (jumpTableInfo->isEmpty() == false) { need to clean the jump table, but
+     // there isn't such an interface yet.  alternatively, replace all the other
+     // blocks in the jump table with the entryBlk //}
+
+  } //wrapup
+
+  static MachineDominatorTree *getDominatorTree(AMDGPUCFGStructurizer &pass) {
+    return &pass.getAnalysis<MachineDominatorTree>();
+  }
+
+  static MachinePostDominatorTree*
+  getPostDominatorTree(AMDGPUCFGStructurizer &pass) {
+    return &pass.getAnalysis<MachinePostDominatorTree>();
+  }
+
+  static MachineLoopInfo *getLoopInfo(AMDGPUCFGStructurizer &pass) {
+    return &pass.getAnalysis<MachineLoopInfo>();
+  }
+}; // template class CFGStructTraits
+} //end of namespace llvm
+
+// createAMDGPUCFGPreparationPass- Returns a pass
+FunctionPass *llvm::createAMDGPUCFGPreparationPass(TargetMachine &tm
+                                                 ) {
+  return new AMDGPUCFGPrepare(tm );
+}
+
+bool AMDGPUCFGPrepare::runOnMachineFunction(MachineFunction &func) {
+  return llvmCFGStruct::CFGStructurizer<AMDGPUCFGStructurizer>().prepare(func,
+                                                                        *this,
+                                                                        TRI);
+}
+
+// createAMDGPUCFGStructurizerPass- Returns a pass
+FunctionPass *llvm::createAMDGPUCFGStructurizerPass(TargetMachine &tm
+                                                  ) {
+  return new AMDGPUCFGPerform(tm );
+}
+
+bool AMDGPUCFGPerform::runOnMachineFunction(MachineFunction &func) {
+  return llvmCFGStruct::CFGStructurizer<AMDGPUCFGStructurizer>().run(func,
+                                                                    *this,
+                                                                    TRI);
+}
diff --git a/lib/Target/R600/AMDILDevice.cpp b/lib/Target/R600/AMDILDevice.cpp
new file mode 100644
index 0000000000..eec5059de2
--- /dev/null
+++ b/lib/Target/R600/AMDILDevice.cpp
@@ -0,0 +1,124 @@
+//===-- AMDILDevice.cpp - Base class for AMDIL Devices --------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//==-----------------------------------------------------------------------===//
+#include "AMDILDevice.h"
+#include "AMDGPUSubtarget.h"
+
+using namespace llvm;
+// Default implementation for all of the classes.
+AMDGPUDevice::AMDGPUDevice(AMDGPUSubtarget *ST) : mSTM(ST) {
+  mHWBits.resize(AMDGPUDeviceInfo::MaxNumberCapabilities);
+  mSWBits.resize(AMDGPUDeviceInfo::MaxNumberCapabilities);
+  setCaps();
+  DeviceFlag = OCL_DEVICE_ALL;
+}
+
+AMDGPUDevice::~AMDGPUDevice() {
+    mHWBits.clear();
+    mSWBits.clear();
+}
+
+size_t AMDGPUDevice::getMaxGDSSize() const {
+  return 0;
+}
+
+uint32_t 
+AMDGPUDevice::getDeviceFlag() const {
+  return DeviceFlag;
+}
+
+size_t AMDGPUDevice::getMaxNumCBs() const {
+  if (usesHardware(AMDGPUDeviceInfo::ConstantMem)) {
+    return HW_MAX_NUM_CB;
+  }
+
+  return 0;
+}
+
+size_t AMDGPUDevice::getMaxCBSize() const {
+  if (usesHardware(AMDGPUDeviceInfo::ConstantMem)) {
+    return MAX_CB_SIZE;
+  }
+
+  return 0;
+}
+
+size_t AMDGPUDevice::getMaxScratchSize() const {
+  return 65536;
+}
+
+uint32_t AMDGPUDevice::getStackAlignment() const {
+  return 16;
+}
+
+void AMDGPUDevice::setCaps() {
+  mSWBits.set(AMDGPUDeviceInfo::HalfOps);
+  mSWBits.set(AMDGPUDeviceInfo::ByteOps);
+  mSWBits.set(AMDGPUDeviceInfo::ShortOps);
+  mSWBits.set(AMDGPUDeviceInfo::HW64BitDivMod);
+  if (mSTM->isOverride(AMDGPUDeviceInfo::NoInline)) {
+    mSWBits.set(AMDGPUDeviceInfo::NoInline);
+  }
+  if (mSTM->isOverride(AMDGPUDeviceInfo::MacroDB)) {
+    mSWBits.set(AMDGPUDeviceInfo::MacroDB);
+  }
+  if (mSTM->isOverride(AMDGPUDeviceInfo::Debug)) {
+    mSWBits.set(AMDGPUDeviceInfo::ConstantMem);
+  } else {
+    mHWBits.set(AMDGPUDeviceInfo::ConstantMem);
+  }
+  if (mSTM->isOverride(AMDGPUDeviceInfo::Debug)) {
+    mSWBits.set(AMDGPUDeviceInfo::PrivateMem);
+  } else {
+    mHWBits.set(AMDGPUDeviceInfo::PrivateMem);
+  }
+  if (mSTM->isOverride(AMDGPUDeviceInfo::BarrierDetect)) {
+    mSWBits.set(AMDGPUDeviceInfo::BarrierDetect);
+  }
+  mSWBits.set(AMDGPUDeviceInfo::ByteLDSOps);
+  mSWBits.set(AMDGPUDeviceInfo::LongOps);
+}
+
+AMDGPUDeviceInfo::ExecutionMode
+AMDGPUDevice::getExecutionMode(AMDGPUDeviceInfo::Caps Caps) const {
+  if (mHWBits[Caps]) {
+    assert(!mSWBits[Caps] && "Cannot set both SW and HW caps");
+    return AMDGPUDeviceInfo::Hardware;
+  }
+
+  if (mSWBits[Caps]) {
+    assert(!mHWBits[Caps] && "Cannot set both SW and HW caps");
+    return AMDGPUDeviceInfo::Software;
+  }
+
+  return AMDGPUDeviceInfo::Unsupported;
+
+}
+
+bool AMDGPUDevice::isSupported(AMDGPUDeviceInfo::Caps Mode) const {
+  return getExecutionMode(Mode) != AMDGPUDeviceInfo::Unsupported;
+}
+
+bool AMDGPUDevice::usesHardware(AMDGPUDeviceInfo::Caps Mode) const {
+  return getExecutionMode(Mode) == AMDGPUDeviceInfo::Hardware;
+}
+
+bool AMDGPUDevice::usesSoftware(AMDGPUDeviceInfo::Caps Mode) const {
+  return getExecutionMode(Mode) == AMDGPUDeviceInfo::Software;
+}
+
+std::string
+AMDGPUDevice::getDataLayout() const {
+    return std::string("e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16"
+      "-i32:32:32-i64:64:64-f32:32:32-f64:64:64-f80:32:32"
+      "-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64"
+      "-v96:128:128-v128:128:128-v192:256:256-v256:256:256"
+      "-v512:512:512-v1024:1024:1024-v2048:2048:2048"
+      "-n8:16:32:64");
+}
diff --git a/lib/Target/R600/AMDILDevice.h b/lib/Target/R600/AMDILDevice.h
new file mode 100644
index 0000000000..b9a15609df
--- /dev/null
+++ b/lib/Target/R600/AMDILDevice.h
@@ -0,0 +1,117 @@
+//===---- AMDILDevice.h - Define Device Data for AMDGPU -----*- C++ -*------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Interface for the subtarget data classes.
+//
+/// This file will define the interface that each generation needs to
+/// implement in order to correctly answer queries on the capabilities of the
+/// specific hardware.
+//===----------------------------------------------------------------------===//
+#ifndef AMDILDEVICEIMPL_H
+#define AMDILDEVICEIMPL_H
+#include "AMDIL.h"
+#include "llvm/ADT/BitVector.h"
+
+namespace llvm {
+  class AMDGPUSubtarget;
+  class MCStreamer;
+//===----------------------------------------------------------------------===//
+// Interface for data that is specific to a single device
+//===----------------------------------------------------------------------===//
+class AMDGPUDevice {
+public:
+  AMDGPUDevice(AMDGPUSubtarget *ST);
+  virtual ~AMDGPUDevice();
+
+  // Enum values for the various memory types.
+  enum {
+    RAW_UAV_ID   = 0,
+    ARENA_UAV_ID = 1,
+    LDS_ID       = 2,
+    GDS_ID       = 3,
+    SCRATCH_ID   = 4,
+    CONSTANT_ID  = 5,
+    GLOBAL_ID    = 6,
+    MAX_IDS      = 7
+  } IO_TYPE_IDS;
+
+  /// \returns The max LDS size that the hardware supports.  Size is in
+  /// bytes.
+  virtual size_t getMaxLDSSize() const = 0;
+
+  /// \returns The max GDS size that the hardware supports if the GDS is
+  /// supported by the hardware.  Size is in bytes.
+  virtual size_t getMaxGDSSize() const;
+
+  /// \returns The max number of hardware constant address spaces that
+  /// are supported by this device.
+  virtual size_t getMaxNumCBs() const;
+
+  /// \returns The max number of bytes a single hardware constant buffer
+  /// can support.  Size is in bytes.
+  virtual size_t getMaxCBSize() const;
+
+  /// \returns The max number of bytes allowed by the hardware scratch
+  /// buffer.  Size is in bytes.
+  virtual size_t getMaxScratchSize() const;
+
+  /// \brief Get the flag that corresponds to the device.
+  virtual uint32_t getDeviceFlag() const;
+
+  /// \returns The number of work-items that exist in a single hardware
+  /// wavefront.
+  virtual size_t getWavefrontSize() const = 0;
+
+  /// \brief Get the generational name of this specific device.
+  virtual uint32_t getGeneration() const = 0;
+
+  /// \brief Get the stack alignment of this specific device.
+  virtual uint32_t getStackAlignment() const;
+
+  /// \brief Get the resource ID for this specific device.
+  virtual uint32_t getResourceID(uint32_t DeviceID) const = 0;
+
+  /// \brief Get the max number of UAV's for this device.
+  virtual uint32_t getMaxNumUAVs() const = 0;
+
+
+  // API utilizing more detailed capabilities of each family of
+  // cards. If a capability is supported, then either usesHardware or
+  // usesSoftware returned true.  If usesHardware returned true, then
+  // usesSoftware must return false for the same capability.  Hardware
+  // execution means that the feature is done natively by the hardware
+  // and is not emulated by the softare.  Software execution means
+  // that the feature could be done in the hardware, but there is
+  // software that emulates it with possibly using the hardware for
+  // support since the hardware does not fully comply with OpenCL
+  // specs.
+
+  bool isSupported(AMDGPUDeviceInfo::Caps Mode) const;
+  bool usesHardware(AMDGPUDeviceInfo::Caps Mode) const;
+  bool usesSoftware(AMDGPUDeviceInfo::Caps Mode) const;
+  virtual std::string getDataLayout() const;
+  static const unsigned int MAX_LDS_SIZE_700 = 16384;
+  static const unsigned int MAX_LDS_SIZE_800 = 32768;
+  static const unsigned int WavefrontSize = 64;
+  static const unsigned int HalfWavefrontSize = 32;
+  static const unsigned int QuarterWavefrontSize = 16;
+protected:
+  virtual void setCaps();
+  llvm::BitVector mHWBits;
+  llvm::BitVector mSWBits;
+  AMDGPUSubtarget *mSTM;
+  uint32_t DeviceFlag;
+private:
+  AMDGPUDeviceInfo::ExecutionMode
+  getExecutionMode(AMDGPUDeviceInfo::Caps Caps) const;
+};
+
+} // namespace llvm
+#endif // AMDILDEVICEIMPL_H
diff --git a/lib/Target/R600/AMDILDeviceInfo.cpp b/lib/Target/R600/AMDILDeviceInfo.cpp
new file mode 100644
index 0000000000..9605fbe633
--- /dev/null
+++ b/lib/Target/R600/AMDILDeviceInfo.cpp
@@ -0,0 +1,94 @@
+//===-- AMDILDeviceInfo.cpp - AMDILDeviceInfo class -----------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Function that creates DeviceInfo from a device name and other information.
+//
+//==-----------------------------------------------------------------------===//
+#include "AMDILDevices.h"
+#include "AMDGPUSubtarget.h"
+
+using namespace llvm;
+namespace llvm {
+namespace AMDGPUDeviceInfo {
+
+AMDGPUDevice* getDeviceFromName(const std::string &deviceName,
+                                AMDGPUSubtarget *ptr,
+                                bool is64bit, bool is64on32bit) {
+  if (deviceName.c_str()[2] == '7') {
+    switch (deviceName.c_str()[3]) {
+    case '1':
+      return new AMDGPU710Device(ptr);
+    case '7':
+      return new AMDGPU770Device(ptr);
+    default:
+      return new AMDGPU7XXDevice(ptr);
+    }
+  } else if (deviceName == "cypress") {
+#if DEBUG
+    assert(!is64bit && "This device does not support 64bit pointers!");
+    assert(!is64on32bit && "This device does not support 64bit"
+          " on 32bit pointers!");
+#endif
+    return new AMDGPUCypressDevice(ptr);
+  } else if (deviceName == "juniper") {
+#if DEBUG
+    assert(!is64bit && "This device does not support 64bit pointers!");
+    assert(!is64on32bit && "This device does not support 64bit"
+          " on 32bit pointers!");
+#endif
+    return new AMDGPUEvergreenDevice(ptr);
+  } else if (deviceName == "redwood") {
+#if DEBUG
+    assert(!is64bit && "This device does not support 64bit pointers!");
+    assert(!is64on32bit && "This device does not support 64bit"
+          " on 32bit pointers!");
+#endif
+    return new AMDGPURedwoodDevice(ptr);
+  } else if (deviceName == "cedar") {
+#if DEBUG
+    assert(!is64bit && "This device does not support 64bit pointers!");
+    assert(!is64on32bit && "This device does not support 64bit"
+          " on 32bit pointers!");
+#endif
+    return new AMDGPUCedarDevice(ptr);
+  } else if (deviceName == "barts" || deviceName == "turks") {
+#if DEBUG
+    assert(!is64bit && "This device does not support 64bit pointers!");
+    assert(!is64on32bit && "This device does not support 64bit"
+          " on 32bit pointers!");
+#endif
+    return new AMDGPUNIDevice(ptr);
+  } else if (deviceName == "cayman") {
+#if DEBUG
+    assert(!is64bit && "This device does not support 64bit pointers!");
+    assert(!is64on32bit && "This device does not support 64bit"
+          " on 32bit pointers!");
+#endif
+    return new AMDGPUCaymanDevice(ptr);
+  } else if (deviceName == "caicos") {
+#if DEBUG
+    assert(!is64bit && "This device does not support 64bit pointers!");
+    assert(!is64on32bit && "This device does not support 64bit"
+          " on 32bit pointers!");
+#endif
+    return new AMDGPUNIDevice(ptr);
+  } else if (deviceName == "SI") {
+    return new AMDGPUSIDevice(ptr);
+  } else {
+#if DEBUG
+    assert(!is64bit && "This device does not support 64bit pointers!");
+    assert(!is64on32bit && "This device does not support 64bit"
+          " on 32bit pointers!");
+#endif
+    return new AMDGPU7XXDevice(ptr);
+  }
+}
+} // End namespace AMDGPUDeviceInfo
+} // End namespace llvm
diff --git a/lib/Target/R600/AMDILDeviceInfo.h b/lib/Target/R600/AMDILDeviceInfo.h
new file mode 100644
index 0000000000..4b2c3a53c7
--- /dev/null
+++ b/lib/Target/R600/AMDILDeviceInfo.h
@@ -0,0 +1,88 @@
+//===-- AMDILDeviceInfo.h - Constants for describing devices --------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//==-----------------------------------------------------------------------===//
+#ifndef AMDILDEVICEINFO_H
+#define AMDILDEVICEINFO_H
+
+
+#include <string>
+
+namespace llvm {
+  class AMDGPUDevice;
+  class AMDGPUSubtarget;
+  namespace AMDGPUDeviceInfo {
+    /// Each Capabilities can be executed using a hardware instruction,
+    /// emulated with a sequence of software instructions, or not
+    /// supported at all.
+    enum ExecutionMode {
+      Unsupported = 0, ///< Unsupported feature on the card(Default value)
+       /// This is the execution mode that is set if the feature is emulated in
+       /// software.
+      Software,
+      /// This execution mode is set if the feature exists natively in hardware
+      Hardware
+    };
+
+    enum Caps {
+      HalfOps          = 0x1,  ///< Half float is supported or not.
+      DoubleOps        = 0x2,  ///< Double is supported or not.
+      ByteOps          = 0x3,  ///< Byte(char) is support or not.
+      ShortOps         = 0x4,  ///< Short is supported or not.
+      LongOps          = 0x5,  ///< Long is supported or not.
+      Images           = 0x6,  ///< Images are supported or not.
+      ByteStores       = 0x7,  ///< ByteStores available(!HD4XXX).
+      ConstantMem      = 0x8,  ///< Constant/CB memory.
+      LocalMem         = 0x9,  ///< Local/LDS memory.
+      PrivateMem       = 0xA,  ///< Scratch/Private/Stack memory.
+      RegionMem        = 0xB,  ///< OCL GDS Memory Extension.
+      FMA              = 0xC,  ///< Use HW FMA or SW FMA.
+      ArenaSegment     = 0xD,  ///< Use for Arena UAV per pointer 12-1023.
+      MultiUAV         = 0xE,  ///< Use for UAV per Pointer 0-7.
+      Reserved0        = 0xF,  ///< ReservedFlag
+      NoAlias          = 0x10, ///< Cached loads.
+      Signed24BitOps   = 0x11, ///< Peephole Optimization.
+      /// Debug mode implies that no hardware features or optimizations
+      /// are performned and that all memory access go through a single
+      /// uav(Arena on HD5XXX/HD6XXX and Raw on HD4XXX).
+      Debug            = 0x12,
+      CachedMem        = 0x13, ///< Cached mem is available or not.
+      BarrierDetect    = 0x14, ///< Detect duplicate barriers.
+      Reserved1        = 0x15, ///< Reserved flag
+      ByteLDSOps       = 0x16, ///< Flag to specify if byte LDS ops are available.
+      ArenaVectors     = 0x17, ///< Flag to specify if vector loads from arena work.
+      TmrReg           = 0x18, ///< Flag to specify if Tmr register is supported.
+      NoInline         = 0x19, ///< Flag to specify that no inlining should occur.
+      MacroDB          = 0x1A, ///< Flag to specify that backend handles macrodb.
+      HW64BitDivMod    = 0x1B, ///< Flag for backend to generate 64bit div/mod.
+      ArenaUAV         = 0x1C, ///< Flag to specify that arena uav is supported.
+      PrivateUAV       = 0x1D, ///< Flag to specify that private memory uses uav's.
+      /// If more capabilities are required, then
+      /// this number needs to be increased.
+      /// All capabilities must come before this
+      /// number.
+      MaxNumberCapabilities = 0x20
+    };
+    /// These have to be in order with the older generations
+    /// having the lower number enumerations.
+    enum Generation {
+      HD4XXX = 0, ///< 7XX based devices.
+      HD5XXX, ///< Evergreen based devices.
+      HD6XXX, ///< NI/Evergreen+ based devices.
+      HD7XXX, ///< Southern Islands based devices.
+      HDTEST, ///< Experimental feature testing device.
+      HDNUMGEN
+    };
+
+
+  AMDGPUDevice*
+    getDeviceFromName(const std::string &name, AMDGPUSubtarget *ptr,
+                      bool is64bit = false, bool is64on32bit = false);
+  } // namespace AMDILDeviceInfo
+} // namespace llvm
+#endif // AMDILDEVICEINFO_H
diff --git a/lib/Target/R600/AMDILDevices.h b/lib/Target/R600/AMDILDevices.h
new file mode 100644
index 0000000000..636fa6d359
--- /dev/null
+++ b/lib/Target/R600/AMDILDevices.h
@@ -0,0 +1,19 @@
+//===-- AMDILDevices.h - Consolidate AMDIL Device headers -----------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//==-----------------------------------------------------------------------===//
+#ifndef AMDIL_DEVICES_H
+#define AMDIL_DEVICES_H
+// Include all of the device specific header files
+#include "AMDIL7XXDevice.h"
+#include "AMDILDevice.h"
+#include "AMDILEvergreenDevice.h"
+#include "AMDILNIDevice.h"
+#include "AMDILSIDevice.h"
+
+#endif // AMDIL_DEVICES_H
diff --git a/lib/Target/R600/AMDILEvergreenDevice.cpp b/lib/Target/R600/AMDILEvergreenDevice.cpp
new file mode 100644
index 0000000000..c5213a0410
--- /dev/null
+++ b/lib/Target/R600/AMDILEvergreenDevice.cpp
@@ -0,0 +1,169 @@
+//===-- AMDILEvergreenDevice.cpp - Device Info for Evergreen --------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//==-----------------------------------------------------------------------===//
+#include "AMDILEvergreenDevice.h"
+
+using namespace llvm;
+
+AMDGPUEvergreenDevice::AMDGPUEvergreenDevice(AMDGPUSubtarget *ST)
+: AMDGPUDevice(ST) {
+  setCaps();
+  std::string name = ST->getDeviceName();
+  if (name == "cedar") {
+    DeviceFlag = OCL_DEVICE_CEDAR;
+  } else if (name == "redwood") {
+    DeviceFlag = OCL_DEVICE_REDWOOD;
+  } else if (name == "cypress") {
+    DeviceFlag = OCL_DEVICE_CYPRESS;
+  } else {
+    DeviceFlag = OCL_DEVICE_JUNIPER;
+  }
+}
+
+AMDGPUEvergreenDevice::~AMDGPUEvergreenDevice() {
+}
+
+size_t AMDGPUEvergreenDevice::getMaxLDSSize() const {
+  if (usesHardware(AMDGPUDeviceInfo::LocalMem)) {
+    return MAX_LDS_SIZE_800;
+  } else {
+    return 0;
+  }
+}
+size_t AMDGPUEvergreenDevice::getMaxGDSSize() const {
+  if (usesHardware(AMDGPUDeviceInfo::RegionMem)) {
+    return MAX_LDS_SIZE_800;
+  } else {
+    return 0;
+  }
+}
+uint32_t AMDGPUEvergreenDevice::getMaxNumUAVs() const {
+  return 12;
+}
+
+uint32_t AMDGPUEvergreenDevice::getResourceID(uint32_t id) const {
+  switch(id) {
+  default:
+    assert(0 && "ID type passed in is unknown!");
+    break;
+  case CONSTANT_ID:
+  case RAW_UAV_ID:
+    return GLOBAL_RETURN_RAW_UAV_ID;
+  case GLOBAL_ID:
+  case ARENA_UAV_ID:
+    return DEFAULT_ARENA_UAV_ID;
+  case LDS_ID:
+    if (usesHardware(AMDGPUDeviceInfo::LocalMem)) {
+      return DEFAULT_LDS_ID;
+    } else {
+      return DEFAULT_ARENA_UAV_ID;
+    }
+  case GDS_ID:
+    if (usesHardware(AMDGPUDeviceInfo::RegionMem)) {
+      return DEFAULT_GDS_ID;
+    } else {
+      return DEFAULT_ARENA_UAV_ID;
+    }
+  case SCRATCH_ID:
+    if (usesHardware(AMDGPUDeviceInfo::PrivateMem)) {
+      return DEFAULT_SCRATCH_ID;
+    } else {
+      return DEFAULT_ARENA_UAV_ID;
+    }
+  };
+  return 0;
+}
+
+size_t AMDGPUEvergreenDevice::getWavefrontSize() const {
+  return AMDGPUDevice::WavefrontSize;
+}
+
+uint32_t AMDGPUEvergreenDevice::getGeneration() const {
+  return AMDGPUDeviceInfo::HD5XXX;
+}
+
+void AMDGPUEvergreenDevice::setCaps() {
+  mSWBits.set(AMDGPUDeviceInfo::ArenaSegment);
+  mHWBits.set(AMDGPUDeviceInfo::ArenaUAV);
+  mHWBits.set(AMDGPUDeviceInfo::HW64BitDivMod);
+  mSWBits.reset(AMDGPUDeviceInfo::HW64BitDivMod);
+  mSWBits.set(AMDGPUDeviceInfo::Signed24BitOps);
+  if (mSTM->isOverride(AMDGPUDeviceInfo::ByteStores)) {
+    mHWBits.set(AMDGPUDeviceInfo::ByteStores);
+  }
+  if (mSTM->isOverride(AMDGPUDeviceInfo::Debug)) {
+    mSWBits.set(AMDGPUDeviceInfo::LocalMem);
+    mSWBits.set(AMDGPUDeviceInfo::RegionMem);
+  } else {
+    mHWBits.set(AMDGPUDeviceInfo::LocalMem);
+    mHWBits.set(AMDGPUDeviceInfo::RegionMem);
+  }
+  mHWBits.set(AMDGPUDeviceInfo::Images);
+  if (mSTM->isOverride(AMDGPUDeviceInfo::NoAlias)) {
+    mHWBits.set(AMDGPUDeviceInfo::NoAlias);
+  }
+  mHWBits.set(AMDGPUDeviceInfo::CachedMem);
+  if (mSTM->isOverride(AMDGPUDeviceInfo::MultiUAV)) {
+    mHWBits.set(AMDGPUDeviceInfo::MultiUAV);
+  }
+  mHWBits.set(AMDGPUDeviceInfo::ByteLDSOps);
+  mSWBits.reset(AMDGPUDeviceInfo::ByteLDSOps);
+  mHWBits.set(AMDGPUDeviceInfo::ArenaVectors);
+  mHWBits.set(AMDGPUDeviceInfo::LongOps);
+  mSWBits.reset(AMDGPUDeviceInfo::LongOps);
+  mHWBits.set(AMDGPUDeviceInfo::TmrReg);
+}
+
+AMDGPUCypressDevice::AMDGPUCypressDevice(AMDGPUSubtarget *ST)
+  : AMDGPUEvergreenDevice(ST) {
+  setCaps();
+}
+
+AMDGPUCypressDevice::~AMDGPUCypressDevice() {
+}
+
+void AMDGPUCypressDevice::setCaps() {
+  if (mSTM->isOverride(AMDGPUDeviceInfo::DoubleOps)) {
+    mHWBits.set(AMDGPUDeviceInfo::DoubleOps);
+    mHWBits.set(AMDGPUDeviceInfo::FMA);
+  }
+}
+
+
+AMDGPUCedarDevice::AMDGPUCedarDevice(AMDGPUSubtarget *ST)
+  : AMDGPUEvergreenDevice(ST) {
+  setCaps();
+}
+
+AMDGPUCedarDevice::~AMDGPUCedarDevice() {
+}
+
+void AMDGPUCedarDevice::setCaps() {
+  mSWBits.set(AMDGPUDeviceInfo::FMA);
+}
+
+size_t AMDGPUCedarDevice::getWavefrontSize() const {
+  return AMDGPUDevice::QuarterWavefrontSize;
+}
+
+AMDGPURedwoodDevice::AMDGPURedwoodDevice(AMDGPUSubtarget *ST)
+  : AMDGPUEvergreenDevice(ST) {
+  setCaps();
+}
+
+AMDGPURedwoodDevice::~AMDGPURedwoodDevice() {
+}
+
+void AMDGPURedwoodDevice::setCaps() {
+  mSWBits.set(AMDGPUDeviceInfo::FMA);
+}
+
+size_t AMDGPURedwoodDevice::getWavefrontSize() const {
+  return AMDGPUDevice::HalfWavefrontSize;
+}
diff --git a/lib/Target/R600/AMDILEvergreenDevice.h b/lib/Target/R600/AMDILEvergreenDevice.h
new file mode 100644
index 0000000000..6dc2deb9ed
--- /dev/null
+++ b/lib/Target/R600/AMDILEvergreenDevice.h
@@ -0,0 +1,93 @@
+//==- AMDILEvergreenDevice.h - Define Evergreen Device for AMDIL -*- C++ -*--=//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Interface for the subtarget data classes.
+///
+/// This file will define the interface that each generation needs to
+/// implement in order to correctly answer queries on the capabilities of the
+/// specific hardware.
+//===----------------------------------------------------------------------===//
+#ifndef AMDILEVERGREENDEVICE_H
+#define AMDILEVERGREENDEVICE_H
+#include "AMDILDevice.h"
+#include "AMDGPUSubtarget.h"
+
+namespace llvm {
+  class AMDGPUSubtarget;
+//===----------------------------------------------------------------------===//
+// Evergreen generation of devices and their respective sub classes
+//===----------------------------------------------------------------------===//
+
+
+/// \brief The AMDGPUEvergreenDevice is the base device class for all of the Evergreen
+/// series of cards.
+///
+/// This class contains information required to differentiate
+/// the Evergreen device from the generic AMDGPUDevice. This device represents
+/// that capabilities of the 'Juniper' cards, also known as the HD57XX.
+class AMDGPUEvergreenDevice : public AMDGPUDevice {
+public:
+  AMDGPUEvergreenDevice(AMDGPUSubtarget *ST);
+  virtual ~AMDGPUEvergreenDevice();
+  virtual size_t getMaxLDSSize() const;
+  virtual size_t getMaxGDSSize() const;
+  virtual size_t getWavefrontSize() const;
+  virtual uint32_t getGeneration() const;
+  virtual uint32_t getMaxNumUAVs() const;
+  virtual uint32_t getResourceID(uint32_t) const;
+protected:
+  virtual void setCaps();
+};
+
+/// The AMDGPUCypressDevice is similiar to the AMDGPUEvergreenDevice, except it has
+/// support for double precision operations. This device is used to represent
+/// both the Cypress and Hemlock cards, which are commercially known as HD58XX
+/// and HD59XX cards.
+class AMDGPUCypressDevice : public AMDGPUEvergreenDevice {
+public:
+  AMDGPUCypressDevice(AMDGPUSubtarget *ST);
+  virtual ~AMDGPUCypressDevice();
+private:
+  virtual void setCaps();
+};
+
+
+/// \brief The AMDGPUCedarDevice is the class that represents all of the 'Cedar' based
+/// devices.
+///
+/// This class differs from the base AMDGPUEvergreenDevice in that the
+/// device is a ~quarter of the 'Juniper'. These are commercially known as the
+/// HD54XX and HD53XX series of cards.
+class AMDGPUCedarDevice : public AMDGPUEvergreenDevice {
+public:
+  AMDGPUCedarDevice(AMDGPUSubtarget *ST);
+  virtual ~AMDGPUCedarDevice();
+  virtual size_t getWavefrontSize() const;
+private:
+  virtual void setCaps();
+};
+
+/// \brief The AMDGPURedwoodDevice is the class the represents all of the 'Redwood' based
+/// devices.
+///
+/// This class differs from the base class, in that these devices are
+/// considered about half of a 'Juniper' device. These are commercially known as
+/// the HD55XX and HD56XX series of cards.
+class AMDGPURedwoodDevice : public AMDGPUEvergreenDevice {
+public:
+  AMDGPURedwoodDevice(AMDGPUSubtarget *ST);
+  virtual ~AMDGPURedwoodDevice();
+  virtual size_t getWavefrontSize() const;
+private:
+  virtual void setCaps();
+};
+  
+} // namespace llvm
+#endif // AMDILEVERGREENDEVICE_H
diff --git a/lib/Target/R600/AMDILFrameLowering.cpp b/lib/Target/R600/AMDILFrameLowering.cpp
new file mode 100644
index 0000000000..9ad495ab48
--- /dev/null
+++ b/lib/Target/R600/AMDILFrameLowering.cpp
@@ -0,0 +1,47 @@
+//===----------------------- AMDILFrameLowering.cpp -----------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Interface to describe a layout of a stack frame on a AMDGPU target
+/// machine.
+//
+//===----------------------------------------------------------------------===//
+#include "AMDILFrameLowering.h"
+#include "llvm/CodeGen/MachineFrameInfo.h"
+
+using namespace llvm;
+AMDGPUFrameLowering::AMDGPUFrameLowering(StackDirection D, unsigned StackAl,
+    int LAO, unsigned TransAl)
+  : TargetFrameLowering(D, StackAl, LAO, TransAl) {
+}
+
+AMDGPUFrameLowering::~AMDGPUFrameLowering() {
+}
+
+int AMDGPUFrameLowering::getFrameIndexOffset(const MachineFunction &MF,
+                                         int FI) const {
+  const MachineFrameInfo *MFI = MF.getFrameInfo();
+  return MFI->getObjectOffset(FI);
+}
+
+const TargetFrameLowering::SpillSlot *
+AMDGPUFrameLowering::getCalleeSavedSpillSlots(unsigned &NumEntries) const {
+  NumEntries = 0;
+  return 0;
+}
+void
+AMDGPUFrameLowering::emitPrologue(MachineFunction &MF) const {
+}
+void
+AMDGPUFrameLowering::emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const {
+}
+bool
+AMDGPUFrameLowering::hasFP(const MachineFunction &MF) const {
+  return false;
+}
diff --git a/lib/Target/R600/AMDILFrameLowering.h b/lib/Target/R600/AMDILFrameLowering.h
new file mode 100644
index 0000000000..51337c3dd2
--- /dev/null
+++ b/lib/Target/R600/AMDILFrameLowering.h
@@ -0,0 +1,40 @@
+//===--------------------- AMDILFrameLowering.h -----------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Interface to describe a layout of a stack frame on a AMDIL target
+/// machine.
+//
+//===----------------------------------------------------------------------===//
+#ifndef AMDILFRAME_LOWERING_H
+#define AMDILFRAME_LOWERING_H
+
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/Target/TargetFrameLowering.h"
+
+namespace llvm {
+
+/// \brief Information about the stack frame layout on the AMDGPU targets.
+///
+/// It holds the direction of the stack growth, the known stack alignment on
+/// entry to each function, and the offset to the locals area.
+/// See TargetFrameInfo for more comments.
+class AMDGPUFrameLowering : public TargetFrameLowering {
+public:
+  AMDGPUFrameLowering(StackDirection D, unsigned StackAl, int LAO,
+                      unsigned TransAl = 1);
+  virtual ~AMDGPUFrameLowering();
+  virtual int getFrameIndexOffset(const MachineFunction &MF, int FI) const;
+  virtual const SpillSlot *getCalleeSavedSpillSlots(unsigned &NumEntries) const;
+  virtual void emitPrologue(MachineFunction &MF) const;
+  virtual void emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const;
+  virtual bool hasFP(const MachineFunction &MF) const;
+};
+} // namespace llvm
+#endif // AMDILFRAME_LOWERING_H
diff --git a/lib/Target/R600/AMDILISelDAGToDAG.cpp b/lib/Target/R600/AMDILISelDAGToDAG.cpp
new file mode 100644
index 0000000000..d15ed393c1
--- /dev/null
+++ b/lib/Target/R600/AMDILISelDAGToDAG.cpp
@@ -0,0 +1,485 @@
+//===-- AMDILISelDAGToDAG.cpp - A dag to dag inst selector for AMDIL ------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Defines an instruction selector for the AMDGPU target.
+//
+//===----------------------------------------------------------------------===//
+#include "AMDGPUInstrInfo.h"
+#include "AMDGPUISelLowering.h" // For AMDGPUISD
+#include "AMDGPURegisterInfo.h"
+#include "AMDILDevices.h"
+#include "R600InstrInfo.h"
+#include "llvm/ADT/ValueMap.h"
+#include "llvm/CodeGen/PseudoSourceValue.h"
+#include "llvm/CodeGen/SelectionDAGISel.h"
+#include "llvm/Support/Compiler.h"
+#include <list>
+#include <queue>
+
+using namespace llvm;
+
+//===----------------------------------------------------------------------===//
+// Instruction Selector Implementation
+//===----------------------------------------------------------------------===//
+
+namespace {
+/// AMDGPU specific code to select AMDGPU machine instructions for
+/// SelectionDAG operations.
+class AMDGPUDAGToDAGISel : public SelectionDAGISel {
+  // Subtarget - Keep a pointer to the AMDGPU Subtarget around so that we can
+  // make the right decision when generating code for different targets.
+  const AMDGPUSubtarget &Subtarget;
+public:
+  AMDGPUDAGToDAGISel(TargetMachine &TM);
+  virtual ~AMDGPUDAGToDAGISel();
+
+  SDNode *Select(SDNode *N);
+  virtual const char *getPassName() const;
+
+private:
+  inline SDValue getSmallIPtrImm(unsigned Imm);
+
+  // Complex pattern selectors
+  bool SelectADDRParam(SDValue Addr, SDValue& R1, SDValue& R2);
+  bool SelectADDR(SDValue N, SDValue &R1, SDValue &R2);
+  bool SelectADDR64(SDValue N, SDValue &R1, SDValue &R2);
+
+  static bool checkType(const Value *ptr, unsigned int addrspace);
+  static const Value *getBasePointerValue(const Value *V);
+
+  static bool isGlobalStore(const StoreSDNode *N);
+  static bool isPrivateStore(const StoreSDNode *N);
+  static bool isLocalStore(const StoreSDNode *N);
+  static bool isRegionStore(const StoreSDNode *N);
+
+  static bool isCPLoad(const LoadSDNode *N);
+  static bool isConstantLoad(const LoadSDNode *N, int cbID);
+  static bool isGlobalLoad(const LoadSDNode *N);
+  static bool isParamLoad(const LoadSDNode *N);
+  static bool isPrivateLoad(const LoadSDNode *N);
+  static bool isLocalLoad(const LoadSDNode *N);
+  static bool isRegionLoad(const LoadSDNode *N);
+
+  bool SelectADDR8BitOffset(SDValue Addr, SDValue& Base, SDValue& Offset);
+  bool SelectADDRReg(SDValue Addr, SDValue& Base, SDValue& Offset);
+  bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset);
+
+  // Include the pieces autogenerated from the target description.
+#include "AMDGPUGenDAGISel.inc"
+};
+}  // end anonymous namespace
+
+/// \brief This pass converts a legalized DAG into a AMDGPU-specific
+// DAG, ready for instruction scheduling.
+FunctionPass *llvm::createAMDGPUISelDag(TargetMachine &TM
+                                       ) {
+  return new AMDGPUDAGToDAGISel(TM);
+}
+
+AMDGPUDAGToDAGISel::AMDGPUDAGToDAGISel(TargetMachine &TM
+                                     )
+  : SelectionDAGISel(TM), Subtarget(TM.getSubtarget<AMDGPUSubtarget>()) {
+}
+
+AMDGPUDAGToDAGISel::~AMDGPUDAGToDAGISel() {
+}
+
+SDValue AMDGPUDAGToDAGISel::getSmallIPtrImm(unsigned int Imm) {
+  return CurDAG->getTargetConstant(Imm, MVT::i32);
+}
+
+bool AMDGPUDAGToDAGISel::SelectADDRParam(
+    SDValue Addr, SDValue& R1, SDValue& R2) {
+
+  if (Addr.getOpcode() == ISD::FrameIndex) {
+    if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
+      R1 = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i32);
+      R2 = CurDAG->getTargetConstant(0, MVT::i32);
+    } else {
+      R1 = Addr;
+      R2 = CurDAG->getTargetConstant(0, MVT::i32);
+    }
+  } else if (Addr.getOpcode() == ISD::ADD) {
+    R1 = Addr.getOperand(0);
+    R2 = Addr.getOperand(1);
+  } else {
+    R1 = Addr;
+    R2 = CurDAG->getTargetConstant(0, MVT::i32);
+  }
+  return true;
+}
+
+bool AMDGPUDAGToDAGISel::SelectADDR(SDValue Addr, SDValue& R1, SDValue& R2) {
+  if (Addr.getOpcode() == ISD::TargetExternalSymbol ||
+      Addr.getOpcode() == ISD::TargetGlobalAddress) {
+    return false;
+  }
+  return SelectADDRParam(Addr, R1, R2);
+}
+
+
+bool AMDGPUDAGToDAGISel::SelectADDR64(SDValue Addr, SDValue& R1, SDValue& R2) {
+  if (Addr.getOpcode() == ISD::TargetExternalSymbol ||
+      Addr.getOpcode() == ISD::TargetGlobalAddress) {
+    return false;
+  }
+
+  if (Addr.getOpcode() == ISD::FrameIndex) {
+    if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
+      R1 = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i64);
+      R2 = CurDAG->getTargetConstant(0, MVT::i64);
+    } else {
+      R1 = Addr;
+      R2 = CurDAG->getTargetConstant(0, MVT::i64);
+    }
+  } else if (Addr.getOpcode() == ISD::ADD) {
+    R1 = Addr.getOperand(0);
+    R2 = Addr.getOperand(1);
+  } else {
+    R1 = Addr;
+    R2 = CurDAG->getTargetConstant(0, MVT::i64);
+  }
+  return true;
+}
+
+SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
+  unsigned int Opc = N->getOpcode();
+  if (N->isMachineOpcode()) {
+    return NULL;   // Already selected.
+  }
+  switch (Opc) {
+  default: break;
+  case ISD::FrameIndex: {
+    if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(N)) {
+      unsigned int FI = FIN->getIndex();
+      EVT OpVT = N->getValueType(0);
+      unsigned int NewOpc = AMDGPU::COPY;
+      SDValue TFI = CurDAG->getTargetFrameIndex(FI, MVT::i32);
+      return CurDAG->SelectNodeTo(N, NewOpc, OpVT, TFI);
+    }
+    break;
+  }
+  case ISD::ConstantFP:
+  case ISD::Constant: {
+    const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
+    // XXX: Custom immediate lowering not implemented yet.  Instead we use
+    // pseudo instructions defined in SIInstructions.td
+    if (ST.device()->getGeneration() > AMDGPUDeviceInfo::HD6XXX) {
+      break;
+    }
+    const R600InstrInfo *TII = static_cast<const R600InstrInfo*>(TM.getInstrInfo());
+
+    uint64_t ImmValue = 0;
+    unsigned ImmReg = AMDGPU::ALU_LITERAL_X;
+
+    if (N->getOpcode() == ISD::ConstantFP) {
+      // XXX: 64-bit Immediates not supported yet
+      assert(N->getValueType(0) != MVT::f64);
+
+      ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N);
+      APFloat Value = C->getValueAPF();
+      float FloatValue = Value.convertToFloat();
+      if (FloatValue == 0.0) {
+        ImmReg = AMDGPU::ZERO;
+      } else if (FloatValue == 0.5) {
+        ImmReg = AMDGPU::HALF;
+      } else if (FloatValue == 1.0) {
+        ImmReg = AMDGPU::ONE;
+      } else {
+        ImmValue = Value.bitcastToAPInt().getZExtValue();
+      }
+    } else {
+      // XXX: 64-bit Immediates not supported yet
+      assert(N->getValueType(0) != MVT::i64);
+
+      ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
+      if (C->getZExtValue() == 0) {
+        ImmReg = AMDGPU::ZERO;
+      } else if (C->getZExtValue() == 1) {
+        ImmReg = AMDGPU::ONE_INT;
+      } else {
+        ImmValue = C->getZExtValue();
+      }
+    }
+
+    for (SDNode::use_iterator Use = N->use_begin(), Next = llvm::next(Use);
+                              Use != SDNode::use_end(); Use = Next) {
+      Next = llvm::next(Use);
+      std::vector<SDValue> Ops;
+      for (unsigned i = 0; i < Use->getNumOperands(); ++i) {
+        Ops.push_back(Use->getOperand(i));
+      }
+
+      if (!Use->isMachineOpcode()) {
+          if (ImmReg == AMDGPU::ALU_LITERAL_X) {
+            // We can only use literal constants (e.g. AMDGPU::ZERO,
+            // AMDGPU::ONE, etc) in machine opcodes.
+            continue;
+          }
+      } else {
+        if (!TII->isALUInstr(Use->getMachineOpcode())) {
+          continue;
+        }
+
+        int ImmIdx = TII->getOperandIdx(Use->getMachineOpcode(), R600Operands::IMM);
+        assert(ImmIdx != -1);
+
+        // subtract one from ImmIdx, because the DST operand is usually index
+        // 0 for MachineInstrs, but we have no DST in the Ops vector.
+        ImmIdx--;
+
+        // Check that we aren't already using an immediate.
+        // XXX: It's possible for an instruction to have more than one
+        // immediate operand, but this is not supported yet.
+        if (ImmReg == AMDGPU::ALU_LITERAL_X) {
+          ConstantSDNode *C = dyn_cast<ConstantSDNode>(Use->getOperand(ImmIdx));
+          assert(C);
+
+          if (C->getZExtValue() != 0) {
+            // This instruction is already using an immediate.
+            continue;
+          }
+
+          // Set the immediate value
+          Ops[ImmIdx] = CurDAG->getTargetConstant(ImmValue, MVT::i32);
+        }
+      }
+      // Set the immediate register
+      Ops[Use.getOperandNo()] = CurDAG->getRegister(ImmReg, MVT::i32);
+
+      CurDAG->UpdateNodeOperands(*Use, Ops.data(), Use->getNumOperands());
+    }
+    break;
+  }
+  }
+  return SelectCode(N);
+}
+
+bool AMDGPUDAGToDAGISel::checkType(const Value *ptr, unsigned int addrspace) {
+  if (!ptr) {
+    return false;
+  }
+  Type *ptrType = ptr->getType();
+  return dyn_cast<PointerType>(ptrType)->getAddressSpace() == addrspace;
+}
+
+const Value * AMDGPUDAGToDAGISel::getBasePointerValue(const Value *V) {
+  if (!V) {
+    return NULL;
+  }
+  const Value *ret = NULL;
+  ValueMap<const Value *, bool> ValueBitMap;
+  std::queue<const Value *, std::list<const Value *> > ValueQueue;
+  ValueQueue.push(V);
+  while (!ValueQueue.empty()) {
+    V = ValueQueue.front();
+    if (ValueBitMap.find(V) == ValueBitMap.end()) {
+      ValueBitMap[V] = true;
+      if (dyn_cast<Argument>(V) && dyn_cast<PointerType>(V->getType())) {
+        ret = V;
+        break;
+      } else if (dyn_cast<GlobalVariable>(V)) {
+        ret = V;
+        break;
+      } else if (dyn_cast<Constant>(V)) {
+        const ConstantExpr *CE = dyn_cast<ConstantExpr>(V);
+        if (CE) {
+          ValueQueue.push(CE->getOperand(0));
+        }
+      } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
+        ret = AI;
+        break;
+      } else if (const Instruction *I = dyn_cast<Instruction>(V)) {
+        uint32_t numOps = I->getNumOperands();
+        for (uint32_t x = 0; x < numOps; ++x) {
+          ValueQueue.push(I->getOperand(x));
+        }
+      } else {
+        assert(!"Found a Value that we didn't know how to handle!");
+      }
+    }
+    ValueQueue.pop();
+  }
+  return ret;
+}
+
+bool AMDGPUDAGToDAGISel::isGlobalStore(const StoreSDNode *N) {
+  return checkType(N->getSrcValue(), AMDGPUAS::GLOBAL_ADDRESS);
+}
+
+bool AMDGPUDAGToDAGISel::isPrivateStore(const StoreSDNode *N) {
+  return (!checkType(N->getSrcValue(), AMDGPUAS::LOCAL_ADDRESS)
+          && !checkType(N->getSrcValue(), AMDGPUAS::GLOBAL_ADDRESS)
+          && !checkType(N->getSrcValue(), AMDGPUAS::REGION_ADDRESS));
+}
+
+bool AMDGPUDAGToDAGISel::isLocalStore(const StoreSDNode *N) {
+  return checkType(N->getSrcValue(), AMDGPUAS::LOCAL_ADDRESS);
+}
+
+bool AMDGPUDAGToDAGISel::isRegionStore(const StoreSDNode *N) {
+  return checkType(N->getSrcValue(), AMDGPUAS::REGION_ADDRESS);
+}
+
+bool AMDGPUDAGToDAGISel::isConstantLoad(const LoadSDNode *N, int cbID) {
+  if (checkType(N->getSrcValue(), AMDGPUAS::CONSTANT_ADDRESS)) {
+    return true;
+  }
+  MachineMemOperand *MMO = N->getMemOperand();
+  const Value *V = MMO->getValue();
+  const Value *BV = getBasePointerValue(V);
+  if (MMO
+      && MMO->getValue()
+      && ((V && dyn_cast<GlobalValue>(V))
+          || (BV && dyn_cast<GlobalValue>(
+                        getBasePointerValue(MMO->getValue()))))) {
+    return checkType(N->getSrcValue(), AMDGPUAS::PRIVATE_ADDRESS);
+  } else {
+    return false;
+  }
+}
+
+bool AMDGPUDAGToDAGISel::isGlobalLoad(const LoadSDNode *N) {
+  return checkType(N->getSrcValue(), AMDGPUAS::GLOBAL_ADDRESS);
+}
+
+bool AMDGPUDAGToDAGISel::isParamLoad(const LoadSDNode *N) {
+  return checkType(N->getSrcValue(), AMDGPUAS::PARAM_I_ADDRESS);
+}
+
+bool AMDGPUDAGToDAGISel::isLocalLoad(const  LoadSDNode *N) {
+  return checkType(N->getSrcValue(), AMDGPUAS::LOCAL_ADDRESS);
+}
+
+bool AMDGPUDAGToDAGISel::isRegionLoad(const  LoadSDNode *N) {
+  return checkType(N->getSrcValue(), AMDGPUAS::REGION_ADDRESS);
+}
+
+bool AMDGPUDAGToDAGISel::isCPLoad(const LoadSDNode *N) {
+  MachineMemOperand *MMO = N->getMemOperand();
+  if (checkType(N->getSrcValue(), AMDGPUAS::PRIVATE_ADDRESS)) {
+    if (MMO) {
+      const Value *V = MMO->getValue();
+      const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V);
+      if (PSV && PSV == PseudoSourceValue::getConstantPool()) {
+        return true;
+      }
+    }
+  }
+  return false;
+}
+
+bool AMDGPUDAGToDAGISel::isPrivateLoad(const LoadSDNode *N) {
+  if (checkType(N->getSrcValue(), AMDGPUAS::PRIVATE_ADDRESS)) {
+    // Check to make sure we are not a constant pool load or a constant load
+    // that is marked as a private load
+    if (isCPLoad(N) || isConstantLoad(N, -1)) {
+      return false;
+    }
+  }
+  if (!checkType(N->getSrcValue(), AMDGPUAS::LOCAL_ADDRESS)
+      && !checkType(N->getSrcValue(), AMDGPUAS::GLOBAL_ADDRESS)
+      && !checkType(N->getSrcValue(), AMDGPUAS::REGION_ADDRESS)
+      && !checkType(N->getSrcValue(), AMDGPUAS::CONSTANT_ADDRESS)
+      && !checkType(N->getSrcValue(), AMDGPUAS::PARAM_D_ADDRESS)
+      && !checkType(N->getSrcValue(), AMDGPUAS::PARAM_I_ADDRESS)) {
+    return true;
+  }
+  return false;
+}
+
+const char *AMDGPUDAGToDAGISel::getPassName() const {
+  return "AMDGPU DAG->DAG Pattern Instruction Selection";
+}
+
+#ifdef DEBUGTMP
+#undef INT64_C
+#endif
+#undef DEBUGTMP
+
+///==== AMDGPU Functions ====///
+
+bool AMDGPUDAGToDAGISel::SelectADDR8BitOffset(SDValue Addr, SDValue& Base,
+                                             SDValue& Offset) {
+  if (Addr.getOpcode() == ISD::TargetExternalSymbol ||
+      Addr.getOpcode() == ISD::TargetGlobalAddress) {
+    return false;
+  }
+
+
+  if (Addr.getOpcode() == ISD::ADD) {
+    bool Match = false;
+
+    // Find the base ptr and the offset
+    for (unsigned i = 0; i < Addr.getNumOperands(); i++) {
+      SDValue Arg = Addr.getOperand(i);
+      ConstantSDNode * OffsetNode = dyn_cast<ConstantSDNode>(Arg);
+      // This arg isn't a constant so it must be the base PTR.
+      if (!OffsetNode) {
+        Base = Addr.getOperand(i);
+        continue;
+      }
+      // Check if the constant argument fits in 8-bits.  The offset is in bytes
+      // so we need to convert it to dwords.
+      if (isUInt<8>(OffsetNode->getZExtValue() >> 2)) {
+        Match = true;
+        Offset = CurDAG->getTargetConstant(OffsetNode->getZExtValue() >> 2,
+                                           MVT::i32);
+      }
+    }
+    return Match;
+  }
+
+  // Default case, no offset
+  Base = Addr;
+  Offset = CurDAG->getTargetConstant(0, MVT::i32);
+  return true;
+}
+
+bool AMDGPUDAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base,
+                                           SDValue &Offset) {
+  ConstantSDNode * IMMOffset;
+
+  if (Addr.getOpcode() == ISD::ADD
+      && (IMMOffset = dyn_cast<ConstantSDNode>(Addr.getOperand(1)))
+      && isInt<16>(IMMOffset->getZExtValue())) {
+
+      Base = Addr.getOperand(0);
+      Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), MVT::i32);
+      return true;
+  // If the pointer address is constant, we can move it to the offset field.
+  } else if ((IMMOffset = dyn_cast<ConstantSDNode>(Addr))
+             && isInt<16>(IMMOffset->getZExtValue())) {
+    Base = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
+                                  CurDAG->getEntryNode().getDebugLoc(),
+                                  AMDGPU::ZERO, MVT::i32);
+    Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), MVT::i32);
+    return true;
+  }
+
+  // Default case, no offset
+  Base = Addr;
+  Offset = CurDAG->getTargetConstant(0, MVT::i32);
+  return true;
+}
+
+bool AMDGPUDAGToDAGISel::SelectADDRReg(SDValue Addr, SDValue& Base,
+                                      SDValue& Offset) {
+  if (Addr.getOpcode() == ISD::TargetExternalSymbol ||
+      Addr.getOpcode() == ISD::TargetGlobalAddress  ||
+      Addr.getOpcode() != ISD::ADD) {
+    return false;
+  }
+
+  Base = Addr.getOperand(0);
+  Offset = Addr.getOperand(1);
+
+  return true;
+}
diff --git a/lib/Target/R600/AMDILISelLowering.cpp b/lib/Target/R600/AMDILISelLowering.cpp
new file mode 100644
index 0000000000..6a5d8416dc
--- /dev/null
+++ b/lib/Target/R600/AMDILISelLowering.cpp
@@ -0,0 +1,652 @@
+//===-- AMDILISelLowering.cpp - AMDIL DAG Lowering Implementation ---------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief TargetLowering functions borrowed from AMDIL.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUISelLowering.h"
+#include "AMDGPURegisterInfo.h"
+#include "AMDILDevices.h"
+#include "AMDILIntrinsicInfo.h"
+#include "AMDGPUSubtarget.h"
+#include "llvm/CallingConv.h"
+#include "llvm/CodeGen/MachineFrameInfo.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/PseudoSourceValue.h"
+#include "llvm/CodeGen/SelectionDAG.h"
+#include "llvm/CodeGen/SelectionDAGNodes.h"
+#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Instructions.h"
+#include "llvm/Intrinsics.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/Target/TargetOptions.h"
+
+using namespace llvm;
+//===----------------------------------------------------------------------===//
+// Calling Convention Implementation
+//===----------------------------------------------------------------------===//
+#include "AMDGPUGenCallingConv.inc"
+
+//===----------------------------------------------------------------------===//
+// TargetLowering Implementation Help Functions End
+//===----------------------------------------------------------------------===//
+
+//===----------------------------------------------------------------------===//
+// TargetLowering Class Implementation Begins
+//===----------------------------------------------------------------------===//
+void AMDGPUTargetLowering::InitAMDILLowering() {
+  int types[] = {
+    (int)MVT::i8,
+    (int)MVT::i16,
+    (int)MVT::i32,
+    (int)MVT::f32,
+    (int)MVT::f64,
+    (int)MVT::i64,
+    (int)MVT::v2i8,
+    (int)MVT::v4i8,
+    (int)MVT::v2i16,
+    (int)MVT::v4i16,
+    (int)MVT::v4f32,
+    (int)MVT::v4i32,
+    (int)MVT::v2f32,
+    (int)MVT::v2i32,
+    (int)MVT::v2f64,
+    (int)MVT::v2i64
+  };
+
+  int IntTypes[] = {
+    (int)MVT::i8,
+    (int)MVT::i16,
+    (int)MVT::i32,
+    (int)MVT::i64
+  };
+
+  int FloatTypes[] = {
+    (int)MVT::f32,
+    (int)MVT::f64
+  };
+
+  int VectorTypes[] = {
+    (int)MVT::v2i8,
+    (int)MVT::v4i8,
+    (int)MVT::v2i16,
+    (int)MVT::v4i16,
+    (int)MVT::v4f32,
+    (int)MVT::v4i32,
+    (int)MVT::v2f32,
+    (int)MVT::v2i32,
+    (int)MVT::v2f64,
+    (int)MVT::v2i64
+  };
+  size_t NumTypes = sizeof(types) / sizeof(*types);
+  size_t NumFloatTypes = sizeof(FloatTypes) / sizeof(*FloatTypes);
+  size_t NumIntTypes = sizeof(IntTypes) / sizeof(*IntTypes);
+  size_t NumVectorTypes = sizeof(VectorTypes) / sizeof(*VectorTypes);
+
+  const AMDGPUSubtarget &STM = getTargetMachine().getSubtarget<AMDGPUSubtarget>();
+  // These are the current register classes that are
+  // supported
+
+  for (unsigned int x  = 0; x < NumTypes; ++x) {
+    MVT::SimpleValueType VT = (MVT::SimpleValueType)types[x];
+
+    //FIXME: SIGN_EXTEND_INREG is not meaningful for floating point types
+    // We cannot sextinreg, expand to shifts
+    setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Custom);
+    setOperationAction(ISD::SUBE, VT, Expand);
+    setOperationAction(ISD::SUBC, VT, Expand);
+    setOperationAction(ISD::ADDE, VT, Expand);
+    setOperationAction(ISD::ADDC, VT, Expand);
+    setOperationAction(ISD::BRCOND, VT, Custom);
+    setOperationAction(ISD::BR_JT, VT, Expand);
+    setOperationAction(ISD::BRIND, VT, Expand);
+    // TODO: Implement custom UREM/SREM routines
+    setOperationAction(ISD::SREM, VT, Expand);
+    setOperationAction(ISD::SMUL_LOHI, VT, Expand);
+    setOperationAction(ISD::UMUL_LOHI, VT, Expand);
+    if (VT != MVT::i64 && VT != MVT::v2i64) {
+      setOperationAction(ISD::SDIV, VT, Custom);
+    }
+  }
+  for (unsigned int x = 0; x < NumFloatTypes; ++x) {
+    MVT::SimpleValueType VT = (MVT::SimpleValueType)FloatTypes[x];
+
+    // IL does not have these operations for floating point types
+    setOperationAction(ISD::FP_ROUND_INREG, VT, Expand);
+    setOperationAction(ISD::SETOLT, VT, Expand);
+    setOperationAction(ISD::SETOGE, VT, Expand);
+    setOperationAction(ISD::SETOGT, VT, Expand);
+    setOperationAction(ISD::SETOLE, VT, Expand);
+    setOperationAction(ISD::SETULT, VT, Expand);
+    setOperationAction(ISD::SETUGE, VT, Expand);
+    setOperationAction(ISD::SETUGT, VT, Expand);
+    setOperationAction(ISD::SETULE, VT, Expand);
+  }
+
+  for (unsigned int x = 0; x < NumIntTypes; ++x) {
+    MVT::SimpleValueType VT = (MVT::SimpleValueType)IntTypes[x];
+
+    // GPU also does not have divrem function for signed or unsigned
+    setOperationAction(ISD::SDIVREM, VT, Expand);
+
+    // GPU does not have [S|U]MUL_LOHI functions as a single instruction
+    setOperationAction(ISD::SMUL_LOHI, VT, Expand);
+    setOperationAction(ISD::UMUL_LOHI, VT, Expand);
+
+    // GPU doesn't have a rotl, rotr, or byteswap instruction
+    setOperationAction(ISD::ROTR, VT, Expand);
+    setOperationAction(ISD::BSWAP, VT, Expand);
+
+    // GPU doesn't have any counting operators
+    setOperationAction(ISD::CTPOP, VT, Expand);
+    setOperationAction(ISD::CTTZ, VT, Expand);
+    setOperationAction(ISD::CTLZ, VT, Expand);
+  }
+
+  for (unsigned int ii = 0; ii < NumVectorTypes; ++ii) {
+    MVT::SimpleValueType VT = (MVT::SimpleValueType)VectorTypes[ii];
+
+    setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
+    setOperationAction(ISD::SDIVREM, VT, Expand);
+    setOperationAction(ISD::SMUL_LOHI, VT, Expand);
+    // setOperationAction(ISD::VSETCC, VT, Expand);
+    setOperationAction(ISD::SELECT_CC, VT, Expand);
+
+  }
+  if (STM.device()->isSupported(AMDGPUDeviceInfo::LongOps)) {
+    setOperationAction(ISD::MULHU, MVT::i64, Expand);
+    setOperationAction(ISD::MULHU, MVT::v2i64, Expand);
+    setOperationAction(ISD::MULHS, MVT::i64, Expand);
+    setOperationAction(ISD::MULHS, MVT::v2i64, Expand);
+    setOperationAction(ISD::ADD, MVT::v2i64, Expand);
+    setOperationAction(ISD::SREM, MVT::v2i64, Expand);
+    setOperationAction(ISD::Constant          , MVT::i64  , Legal);
+    setOperationAction(ISD::SDIV, MVT::v2i64, Expand);
+    setOperationAction(ISD::TRUNCATE, MVT::v2i64, Expand);
+    setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Expand);
+    setOperationAction(ISD::ZERO_EXTEND, MVT::v2i64, Expand);
+    setOperationAction(ISD::ANY_EXTEND, MVT::v2i64, Expand);
+  }
+  if (STM.device()->isSupported(AMDGPUDeviceInfo::DoubleOps)) {
+    // we support loading/storing v2f64 but not operations on the type
+    setOperationAction(ISD::FADD, MVT::v2f64, Expand);
+    setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
+    setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
+    setOperationAction(ISD::FP_ROUND_INREG, MVT::v2f64, Expand);
+    setOperationAction(ISD::FP_EXTEND, MVT::v2f64, Expand);
+    setOperationAction(ISD::ConstantFP        , MVT::f64  , Legal);
+    // We want to expand vector conversions into their scalar
+    // counterparts.
+    setOperationAction(ISD::TRUNCATE, MVT::v2f64, Expand);
+    setOperationAction(ISD::SIGN_EXTEND, MVT::v2f64, Expand);
+    setOperationAction(ISD::ZERO_EXTEND, MVT::v2f64, Expand);
+    setOperationAction(ISD::ANY_EXTEND, MVT::v2f64, Expand);
+    setOperationAction(ISD::FABS, MVT::f64, Expand);
+    setOperationAction(ISD::FABS, MVT::v2f64, Expand);
+  }
+  // TODO: Fix the UDIV24 algorithm so it works for these
+  // types correctly. This needs vector comparisons
+  // for this to work correctly.
+  setOperationAction(ISD::UDIV, MVT::v2i8, Expand);
+  setOperationAction(ISD::UDIV, MVT::v4i8, Expand);
+  setOperationAction(ISD::UDIV, MVT::v2i16, Expand);
+  setOperationAction(ISD::UDIV, MVT::v4i16, Expand);
+  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Custom);
+  setOperationAction(ISD::SUBC, MVT::Other, Expand);
+  setOperationAction(ISD::ADDE, MVT::Other, Expand);
+  setOperationAction(ISD::ADDC, MVT::Other, Expand);
+  setOperationAction(ISD::BRCOND, MVT::Other, Custom);
+  setOperationAction(ISD::BR_JT, MVT::Other, Expand);
+  setOperationAction(ISD::BRIND, MVT::Other, Expand);
+  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Expand);
+
+
+  // Use the default implementation.
+  setOperationAction(ISD::ConstantFP        , MVT::f32    , Legal);
+  setOperationAction(ISD::Constant          , MVT::i32    , Legal);
+
+  setSchedulingPreference(Sched::RegPressure);
+  setPow2DivIsCheap(false);
+  setPrefLoopAlignment(16);
+  setSelectIsExpensive(true);
+  setJumpIsExpensive(true);
+
+  maxStoresPerMemcpy  = 4096;
+  maxStoresPerMemmove = 4096;
+  maxStoresPerMemset  = 4096;
+
+}
+
+bool
+AMDGPUTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
+    const CallInst &I, unsigned Intrinsic) const {
+  return false;
+}
+
+// The backend supports 32 and 64 bit floating point immediates
+bool
+AMDGPUTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
+  if (VT.getScalarType().getSimpleVT().SimpleTy == MVT::f32
+      || VT.getScalarType().getSimpleVT().SimpleTy == MVT::f64) {
+    return true;
+  } else {
+    return false;
+  }
+}
+
+bool
+AMDGPUTargetLowering::ShouldShrinkFPConstant(EVT VT) const {
+  if (VT.getScalarType().getSimpleVT().SimpleTy == MVT::f32
+      || VT.getScalarType().getSimpleVT().SimpleTy == MVT::f64) {
+    return false;
+  } else {
+    return true;
+  }
+}
+
+
+// isMaskedValueZeroForTargetNode - Return true if 'Op & Mask' is known to
+// be zero. Op is expected to be a target specific node. Used by DAG
+// combiner.
+
+void
+AMDGPUTargetLowering::computeMaskedBitsForTargetNode(
+    const SDValue Op,
+    APInt &KnownZero,
+    APInt &KnownOne,
+    const SelectionDAG &DAG,
+    unsigned Depth) const {
+  APInt KnownZero2;
+  APInt KnownOne2;
+  KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything
+  switch (Op.getOpcode()) {
+    default: break;
+    case ISD::SELECT_CC:
+             DAG.ComputeMaskedBits(
+                 Op.getOperand(1),
+                 KnownZero,
+                 KnownOne,
+                 Depth + 1
+                 );
+             DAG.ComputeMaskedBits(
+                 Op.getOperand(0),
+                 KnownZero2,
+                 KnownOne2
+                 );
+             assert((KnownZero & KnownOne) == 0
+                 && "Bits known to be one AND zero?");
+             assert((KnownZero2 & KnownOne2) == 0
+                 && "Bits known to be one AND zero?");
+             // Only known if known in both the LHS and RHS
+             KnownOne &= KnownOne2;
+             KnownZero &= KnownZero2;
+             break;
+  };
+}
+
+//===----------------------------------------------------------------------===//
+//                           Other Lowering Hooks
+//===----------------------------------------------------------------------===//
+
+SDValue
+AMDGPUTargetLowering::LowerSDIV(SDValue Op, SelectionDAG &DAG) const {
+  EVT OVT = Op.getValueType();
+  SDValue DST;
+  if (OVT.getScalarType() == MVT::i64) {
+    DST = LowerSDIV64(Op, DAG);
+  } else if (OVT.getScalarType() == MVT::i32) {
+    DST = LowerSDIV32(Op, DAG);
+  } else if (OVT.getScalarType() == MVT::i16
+      || OVT.getScalarType() == MVT::i8) {
+    DST = LowerSDIV24(Op, DAG);
+  } else {
+    DST = SDValue(Op.getNode(), 0);
+  }
+  return DST;
+}
+
+SDValue
+AMDGPUTargetLowering::LowerSREM(SDValue Op, SelectionDAG &DAG) const {
+  EVT OVT = Op.getValueType();
+  SDValue DST;
+  if (OVT.getScalarType() == MVT::i64) {
+    DST = LowerSREM64(Op, DAG);
+  } else if (OVT.getScalarType() == MVT::i32) {
+    DST = LowerSREM32(Op, DAG);
+  } else if (OVT.getScalarType() == MVT::i16) {
+    DST = LowerSREM16(Op, DAG);
+  } else if (OVT.getScalarType() == MVT::i8) {
+    DST = LowerSREM8(Op, DAG);
+  } else {
+    DST = SDValue(Op.getNode(), 0);
+  }
+  return DST;
+}
+
+SDValue
+AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const {
+  SDValue Data = Op.getOperand(0);
+  VTSDNode *BaseType = cast<VTSDNode>(Op.getOperand(1));
+  DebugLoc DL = Op.getDebugLoc();
+  EVT DVT = Data.getValueType();
+  EVT BVT = BaseType->getVT();
+  unsigned baseBits = BVT.getScalarType().getSizeInBits();
+  unsigned srcBits = DVT.isSimple() ? DVT.getScalarType().getSizeInBits() : 1;
+  unsigned shiftBits = srcBits - baseBits;
+  if (srcBits < 32) {
+    // If the op is less than 32 bits, then it needs to extend to 32bits
+    // so it can properly keep the upper bits valid.
+    EVT IVT = genIntType(32, DVT.isVector() ? DVT.getVectorNumElements() : 1);
+    Data = DAG.getNode(ISD::ZERO_EXTEND, DL, IVT, Data);
+    shiftBits = 32 - baseBits;
+    DVT = IVT;
+  }
+  SDValue Shift = DAG.getConstant(shiftBits, DVT);
+  // Shift left by 'Shift' bits.
+  Data = DAG.getNode(ISD::SHL, DL, DVT, Data, Shift);
+  // Signed shift Right by 'Shift' bits.
+  Data = DAG.getNode(ISD::SRA, DL, DVT, Data, Shift);
+  if (srcBits < 32) {
+    // Once the sign extension is done, the op needs to be converted to
+    // its original type.
+    Data = DAG.getSExtOrTrunc(Data, DL, Op.getOperand(0).getValueType());
+  }
+  return Data;
+}
+EVT
+AMDGPUTargetLowering::genIntType(uint32_t size, uint32_t numEle) const {
+  int iSize = (size * numEle);
+  int vEle = (iSize >> ((size == 64) ? 6 : 5));
+  if (!vEle) {
+    vEle = 1;
+  }
+  if (size == 64) {
+    if (vEle == 1) {
+      return EVT(MVT::i64);
+    } else {
+      return EVT(MVT::getVectorVT(MVT::i64, vEle));
+    }
+  } else {
+    if (vEle == 1) {
+      return EVT(MVT::i32);
+    } else {
+      return EVT(MVT::getVectorVT(MVT::i32, vEle));
+    }
+  }
+}
+
+SDValue
+AMDGPUTargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
+  SDValue Chain = Op.getOperand(0);
+  SDValue Cond  = Op.getOperand(1);
+  SDValue Jump  = Op.getOperand(2);
+  SDValue Result;
+  Result = DAG.getNode(
+      AMDGPUISD::BRANCH_COND,
+      Op.getDebugLoc(),
+      Op.getValueType(),
+      Chain, Jump, Cond);
+  return Result;
+}
+
+SDValue
+AMDGPUTargetLowering::LowerSDIV24(SDValue Op, SelectionDAG &DAG) const {
+  DebugLoc DL = Op.getDebugLoc();
+  EVT OVT = Op.getValueType();
+  SDValue LHS = Op.getOperand(0);
+  SDValue RHS = Op.getOperand(1);
+  MVT INTTY;
+  MVT FLTTY;
+  if (!OVT.isVector()) {
+    INTTY = MVT::i32;
+    FLTTY = MVT::f32;
+  } else if (OVT.getVectorNumElements() == 2) {
+    INTTY = MVT::v2i32;
+    FLTTY = MVT::v2f32;
+  } else if (OVT.getVectorNumElements() == 4) {
+    INTTY = MVT::v4i32;
+    FLTTY = MVT::v4f32;
+  }
+  unsigned bitsize = OVT.getScalarType().getSizeInBits();
+  // char|short jq = ia ^ ib;
+  SDValue jq = DAG.getNode(ISD::XOR, DL, OVT, LHS, RHS);
+
+  // jq = jq >> (bitsize - 2)
+  jq = DAG.getNode(ISD::SRA, DL, OVT, jq, DAG.getConstant(bitsize - 2, OVT)); 
+
+  // jq = jq | 0x1
+  jq = DAG.getNode(ISD::OR, DL, OVT, jq, DAG.getConstant(1, OVT));
+
+  // jq = (int)jq
+  jq = DAG.getSExtOrTrunc(jq, DL, INTTY);
+
+  // int ia = (int)LHS;
+  SDValue ia = DAG.getSExtOrTrunc(LHS, DL, INTTY);
+
+  // int ib, (int)RHS;
+  SDValue ib = DAG.getSExtOrTrunc(RHS, DL, INTTY);
+
+  // float fa = (float)ia;
+  SDValue fa = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ia);
+
+  // float fb = (float)ib;
+  SDValue fb = DAG.getNode(ISD::SINT_TO_FP, DL, FLTTY, ib);
+
+  // float fq = native_divide(fa, fb);
+  SDValue fq = DAG.getNode(AMDGPUISD::DIV_INF, DL, FLTTY, fa, fb);
+
+  // fq = trunc(fq);
+  fq = DAG.getNode(ISD::FTRUNC, DL, FLTTY, fq);
+
+  // float fqneg = -fq;
+  SDValue fqneg = DAG.getNode(ISD::FNEG, DL, FLTTY, fq);
+
+  // float fr = mad(fqneg, fb, fa);
+  SDValue fr = DAG.getNode(AMDGPUISD::MAD, DL, FLTTY, fqneg, fb, fa);
+
+  // int iq = (int)fq;
+  SDValue iq = DAG.getNode(ISD::FP_TO_SINT, DL, INTTY, fq);
+
+  // fr = fabs(fr);
+  fr = DAG.getNode(ISD::FABS, DL, FLTTY, fr);
+
+  // fb = fabs(fb);
+  fb = DAG.getNode(ISD::FABS, DL, FLTTY, fb);
+
+  // int cv = fr >= fb;
+  SDValue cv;
+  if (INTTY == MVT::i32) {
+    cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
+  } else {
+    cv = DAG.getSetCC(DL, INTTY, fr, fb, ISD::SETOGE);
+  }
+  // jq = (cv ? jq : 0);
+  jq = DAG.getNode(ISD::SELECT, DL, OVT, cv, jq, 
+      DAG.getConstant(0, OVT));
+  // dst = iq + jq;
+  iq = DAG.getSExtOrTrunc(iq, DL, OVT);
+  iq = DAG.getNode(ISD::ADD, DL, OVT, iq, jq);
+  return iq;
+}
+
+SDValue
+AMDGPUTargetLowering::LowerSDIV32(SDValue Op, SelectionDAG &DAG) const {
+  DebugLoc DL = Op.getDebugLoc();
+  EVT OVT = Op.getValueType();
+  SDValue LHS = Op.getOperand(0);
+  SDValue RHS = Op.getOperand(1);
+  // The LowerSDIV32 function generates equivalent to the following IL.
+  // mov r0, LHS
+  // mov r1, RHS
+  // ilt r10, r0, 0
+  // ilt r11, r1, 0
+  // iadd r0, r0, r10
+  // iadd r1, r1, r11
+  // ixor r0, r0, r10
+  // ixor r1, r1, r11
+  // udiv r0, r0, r1
+  // ixor r10, r10, r11
+  // iadd r0, r0, r10
+  // ixor DST, r0, r10
+
+  // mov r0, LHS
+  SDValue r0 = LHS;
+
+  // mov r1, RHS
+  SDValue r1 = RHS;
+
+  // ilt r10, r0, 0
+  SDValue r10 = DAG.getSelectCC(DL,
+      r0, DAG.getConstant(0, OVT),
+      DAG.getConstant(-1, MVT::i32),
+      DAG.getConstant(0, MVT::i32),
+      ISD::SETLT);
+
+  // ilt r11, r1, 0
+  SDValue r11 = DAG.getSelectCC(DL,
+      r1, DAG.getConstant(0, OVT),
+      DAG.getConstant(-1, MVT::i32),
+      DAG.getConstant(0, MVT::i32),
+      ISD::SETLT);
+
+  // iadd r0, r0, r10
+  r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
+
+  // iadd r1, r1, r11
+  r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
+
+  // ixor r0, r0, r10
+  r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
+
+  // ixor r1, r1, r11
+  r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
+
+  // udiv r0, r0, r1
+  r0 = DAG.getNode(ISD::UDIV, DL, OVT, r0, r1);
+
+  // ixor r10, r10, r11
+  r10 = DAG.getNode(ISD::XOR, DL, OVT, r10, r11);
+
+  // iadd r0, r0, r10
+  r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
+
+  // ixor DST, r0, r10
+  SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10); 
+  return DST;
+}
+
+SDValue
+AMDGPUTargetLowering::LowerSDIV64(SDValue Op, SelectionDAG &DAG) const {
+  return SDValue(Op.getNode(), 0);
+}
+
+SDValue
+AMDGPUTargetLowering::LowerSREM8(SDValue Op, SelectionDAG &DAG) const {
+  DebugLoc DL = Op.getDebugLoc();
+  EVT OVT = Op.getValueType();
+  MVT INTTY = MVT::i32;
+  if (OVT == MVT::v2i8) {
+    INTTY = MVT::v2i32;
+  } else if (OVT == MVT::v4i8) {
+    INTTY = MVT::v4i32;
+  }
+  SDValue LHS = DAG.getSExtOrTrunc(Op.getOperand(0), DL, INTTY);
+  SDValue RHS = DAG.getSExtOrTrunc(Op.getOperand(1), DL, INTTY);
+  LHS = DAG.getNode(ISD::SREM, DL, INTTY, LHS, RHS);
+  LHS = DAG.getSExtOrTrunc(LHS, DL, OVT);
+  return LHS;
+}
+
+SDValue
+AMDGPUTargetLowering::LowerSREM16(SDValue Op, SelectionDAG &DAG) const {
+  DebugLoc DL = Op.getDebugLoc();
+  EVT OVT = Op.getValueType();
+  MVT INTTY = MVT::i32;
+  if (OVT == MVT::v2i16) {
+    INTTY = MVT::v2i32;
+  } else if (OVT == MVT::v4i16) {
+    INTTY = MVT::v4i32;
+  }
+  SDValue LHS = DAG.getSExtOrTrunc(Op.getOperand(0), DL, INTTY);
+  SDValue RHS = DAG.getSExtOrTrunc(Op.getOperand(1), DL, INTTY);
+  LHS = DAG.getNode(ISD::SREM, DL, INTTY, LHS, RHS);
+  LHS = DAG.getSExtOrTrunc(LHS, DL, OVT);
+  return LHS;
+}
+
+SDValue
+AMDGPUTargetLowering::LowerSREM32(SDValue Op, SelectionDAG &DAG) const {
+  DebugLoc DL = Op.getDebugLoc();
+  EVT OVT = Op.getValueType();
+  SDValue LHS = Op.getOperand(0);
+  SDValue RHS = Op.getOperand(1);
+  // The LowerSREM32 function generates equivalent to the following IL.
+  // mov r0, LHS
+  // mov r1, RHS
+  // ilt r10, r0, 0
+  // ilt r11, r1, 0
+  // iadd r0, r0, r10
+  // iadd r1, r1, r11
+  // ixor r0, r0, r10
+  // ixor r1, r1, r11
+  // udiv r20, r0, r1
+  // umul r20, r20, r1
+  // sub r0, r0, r20
+  // iadd r0, r0, r10
+  // ixor DST, r0, r10
+
+  // mov r0, LHS
+  SDValue r0 = LHS;
+
+  // mov r1, RHS
+  SDValue r1 = RHS;
+
+  // ilt r10, r0, 0
+  SDValue r10 = DAG.getSetCC(DL, OVT, r0, DAG.getConstant(0, OVT), ISD::SETLT);
+
+  // ilt r11, r1, 0
+  SDValue r11 = DAG.getSetCC(DL, OVT, r1, DAG.getConstant(0, OVT), ISD::SETLT);
+
+  // iadd r0, r0, r10
+  r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
+
+  // iadd r1, r1, r11
+  r1 = DAG.getNode(ISD::ADD, DL, OVT, r1, r11);
+
+  // ixor r0, r0, r10
+  r0 = DAG.getNode(ISD::XOR, DL, OVT, r0, r10);
+
+  // ixor r1, r1, r11
+  r1 = DAG.getNode(ISD::XOR, DL, OVT, r1, r11);
+
+  // udiv r20, r0, r1
+  SDValue r20 = DAG.getNode(ISD::UREM, DL, OVT, r0, r1);
+
+  // umul r20, r20, r1
+  r20 = DAG.getNode(AMDGPUISD::UMUL, DL, OVT, r20, r1);
+
+  // sub r0, r0, r20
+  r0 = DAG.getNode(ISD::SUB, DL, OVT, r0, r20);
+
+  // iadd r0, r0, r10
+  r0 = DAG.getNode(ISD::ADD, DL, OVT, r0, r10);
+
+  // ixor DST, r0, r10
+  SDValue DST = DAG.getNode(ISD::XOR, DL, OVT, r0, r10); 
+  return DST;
+}
+
+SDValue
+AMDGPUTargetLowering::LowerSREM64(SDValue Op, SelectionDAG &DAG) const {
+  return SDValue(Op.getNode(), 0);
+}
diff --git a/lib/Target/R600/AMDILInstrInfo.td b/lib/Target/R600/AMDILInstrInfo.td
new file mode 100644
index 0000000000..ac6745148e
--- /dev/null
+++ b/lib/Target/R600/AMDILInstrInfo.td
@@ -0,0 +1,273 @@
+//===------------ AMDILInstrInfo.td - AMDIL Target ------*-tablegen-*------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+// This file describes the AMDIL instructions in TableGen format.
+//
+//===----------------------------------------------------------------------===//
+// AMDIL Instruction Predicate Definitions
+// Predicate that is set to true if the hardware supports double precision
+// divide
+def HasHWDDiv                 : Predicate<"Subtarget.device()"
+                           "->getGeneration() > AMDGPUDeviceInfo::HD4XXX && "
+              "Subtarget.device()->usesHardware(AMDGPUDeviceInfo::DoubleOps)">;
+
+// Predicate that is set to true if the hardware supports double, but not double
+// precision divide in hardware
+def HasSWDDiv             : Predicate<"Subtarget.device()"
+                           "->getGeneration() == AMDGPUDeviceInfo::HD4XXX &&"
+              "Subtarget.device()->usesHardware(AMDGPUDeviceInfo::DoubleOps)">;
+
+// Predicate that is set to true if the hardware support 24bit signed
+// math ops. Otherwise a software expansion to 32bit math ops is used instead.
+def HasHWSign24Bit          : Predicate<"Subtarget.device()"
+                            "->getGeneration() > AMDGPUDeviceInfo::HD5XXX">;
+
+// Predicate that is set to true if 64bit operations are supported or not
+def HasHW64Bit              : Predicate<"Subtarget.device()"
+                            "->usesHardware(AMDGPUDeviceInfo::LongOps)">;
+def HasSW64Bit              : Predicate<"Subtarget.device()"
+                            "->usesSoftware(AMDGPUDeviceInfo::LongOps)">;
+
+// Predicate that is set to true if the timer register is supported
+def HasTmrRegister          : Predicate<"Subtarget.device()"
+                            "->isSupported(AMDGPUDeviceInfo::TmrReg)">;
+// Predicate that is true if we are at least evergreen series
+def HasDeviceIDInst         : Predicate<"Subtarget.device()"
+                            "->getGeneration() >= AMDGPUDeviceInfo::HD5XXX">;
+
+// Predicate that is true if we have region address space.
+def hasRegionAS             : Predicate<"Subtarget.device()"
+                            "->usesHardware(AMDGPUDeviceInfo::RegionMem)">;
+
+// Predicate that is false if we don't have region address space.
+def noRegionAS             : Predicate<"!Subtarget.device()"
+                            "->isSupported(AMDGPUDeviceInfo::RegionMem)">;
+
+
+// Predicate that is set to true if 64bit Mul is supported in the IL or not
+def HasHW64Mul              : Predicate<"Subtarget.calVersion()" 
+                                          ">= CAL_VERSION_SC_139"
+                                          "&& Subtarget.device()"
+                                          "->getGeneration() >="
+                                          "AMDGPUDeviceInfo::HD5XXX">;
+def HasSW64Mul              : Predicate<"Subtarget.calVersion()" 
+                                          "< CAL_VERSION_SC_139">;
+// Predicate that is set to true if 64bit Div/Mod is supported in the IL or not
+def HasHW64DivMod           : Predicate<"Subtarget.device()"
+                            "->usesHardware(AMDGPUDeviceInfo::HW64BitDivMod)">;
+def HasSW64DivMod           : Predicate<"Subtarget.device()"
+                            "->usesSoftware(AMDGPUDeviceInfo::HW64BitDivMod)">;
+
+// Predicate that is set to true if 64bit pointer are used.
+def Has64BitPtr             : Predicate<"Subtarget.is64bit()">;
+def Has32BitPtr             : Predicate<"!Subtarget.is64bit()">;
+//===--------------------------------------------------------------------===//
+// Custom Operands
+//===--------------------------------------------------------------------===//
+def brtarget   : Operand<OtherVT>;
+
+//===--------------------------------------------------------------------===//
+// Custom Selection DAG Type Profiles
+//===--------------------------------------------------------------------===//
+//===----------------------------------------------------------------------===//
+// Generic Profile Types
+//===----------------------------------------------------------------------===//
+
+def SDTIL_GenBinaryOp : SDTypeProfile<1, 2, [
+    SDTCisSameAs<0, 1>, SDTCisSameAs<1, 2>
+    ]>;
+def SDTIL_GenTernaryOp : SDTypeProfile<1, 3, [
+    SDTCisSameAs<0, 1>, SDTCisSameAs<1, 2>, SDTCisSameAs<2, 3>
+    ]>;
+def SDTIL_GenVecBuild : SDTypeProfile<1, 1, [
+    SDTCisEltOfVec<1, 0>
+    ]>;
+
+//===----------------------------------------------------------------------===//
+// Flow Control Profile Types
+//===----------------------------------------------------------------------===//
+// Branch instruction where second and third are basic blocks
+def SDTIL_BRCond : SDTypeProfile<0, 2, [
+    SDTCisVT<0, OtherVT>
+    ]>;
+
+//===--------------------------------------------------------------------===//
+// Custom Selection DAG Nodes
+//===--------------------------------------------------------------------===//
+//===----------------------------------------------------------------------===//
+// Flow Control DAG Nodes
+//===----------------------------------------------------------------------===//
+def IL_brcond      : SDNode<"AMDGPUISD::BRANCH_COND", SDTIL_BRCond, [SDNPHasChain]>;
+
+//===----------------------------------------------------------------------===//
+// Call/Return DAG Nodes
+//===----------------------------------------------------------------------===//
+def IL_retflag       : SDNode<"AMDGPUISD::RET_FLAG", SDTNone,
+    [SDNPHasChain, SDNPOptInGlue]>;
+
+//===--------------------------------------------------------------------===//
+// Instructions
+//===--------------------------------------------------------------------===//
+// Floating point math functions
+def IL_div_inf      : SDNode<"AMDGPUISD::DIV_INF", SDTIL_GenBinaryOp>;
+def IL_mad          : SDNode<"AMDGPUISD::MAD", SDTIL_GenTernaryOp>;
+
+//===----------------------------------------------------------------------===//
+// Integer functions
+//===----------------------------------------------------------------------===//
+def IL_umul        : SDNode<"AMDGPUISD::UMUL"    , SDTIntBinOp,
+    [SDNPCommutative, SDNPAssociative]>;
+
+//===--------------------------------------------------------------------===//
+// Custom Pattern DAG Nodes
+//===--------------------------------------------------------------------===//
+def global_store : PatFrag<(ops node:$val, node:$ptr),
+    (store node:$val, node:$ptr), [{
+        return isGlobalStore(dyn_cast<StoreSDNode>(N));
+}]>;
+
+//===----------------------------------------------------------------------===//
+// Load pattern fragments
+//===----------------------------------------------------------------------===//
+// Global address space loads
+def global_load : PatFrag<(ops node:$ptr), (load node:$ptr), [{
+    return isGlobalLoad(dyn_cast<LoadSDNode>(N));
+}]>;
+// Constant address space loads
+def constant_load : PatFrag<(ops node:$ptr), (load node:$ptr), [{
+    return isConstantLoad(dyn_cast<LoadSDNode>(N), -1);
+}]>;
+
+//===----------------------------------------------------------------------===//
+// Complex addressing mode patterns
+//===----------------------------------------------------------------------===//
+def ADDR : ComplexPattern<i32, 2, "SelectADDR", [], []>;
+def ADDRF : ComplexPattern<i32, 2, "SelectADDR", [frameindex], []>;
+def ADDR64 : ComplexPattern<i64, 2, "SelectADDR64", [], []>;
+def ADDR64F : ComplexPattern<i64, 2, "SelectADDR64", [frameindex], []>;
+
+//===----------------------------------------------------------------------===//
+// Instruction format classes
+//===----------------------------------------------------------------------===//
+class ILFormat<dag outs, dag ins, string asmstr, list<dag> pattern>
+: Instruction {
+
+     let Namespace = "AMDGPU";
+     dag OutOperandList = outs;
+     dag InOperandList = ins;
+     let Pattern = pattern;
+     let AsmString = !strconcat(asmstr, "\n");
+     let isPseudo = 1;
+     let Itinerary = NullALU;
+     bit hasIEEEFlag = 0;
+     bit hasZeroOpFlag = 0;
+     let mayLoad = 0;
+     let mayStore = 0;
+     let hasSideEffects = 0;
+}
+
+//===--------------------------------------------------------------------===//
+// Multiclass Instruction formats
+//===--------------------------------------------------------------------===//
+// Multiclass that handles branch instructions
+multiclass BranchConditional<SDNode Op> {
+    def _i32 : ILFormat<(outs),
+  (ins brtarget:$target, GPRI32:$src0),
+        "; i32 Pseudo branch instruction",
+  [(Op bb:$target, GPRI32:$src0)]>;
+    def _f32 : ILFormat<(outs),
+  (ins brtarget:$target, GPRF32:$src0),
+        "; f32 Pseudo branch instruction",
+  [(Op bb:$target, GPRF32:$src0)]>;
+}
+
+// Only scalar types should generate flow control
+multiclass BranchInstr<string name> {
+  def _i32 : ILFormat<(outs), (ins GPRI32:$src),
+      !strconcat(name, " $src"), []>;
+  def _f32 : ILFormat<(outs), (ins GPRF32:$src),
+      !strconcat(name, " $src"), []>;
+}
+// Only scalar types should generate flow control
+multiclass BranchInstr2<string name> {
+  def _i32 : ILFormat<(outs), (ins GPRI32:$src0, GPRI32:$src1),
+      !strconcat(name, " $src0, $src1"), []>;
+  def _f32 : ILFormat<(outs), (ins GPRF32:$src0, GPRF32:$src1),
+      !strconcat(name, " $src0, $src1"), []>;
+}
+
+//===--------------------------------------------------------------------===//
+// Intrinsics support
+//===--------------------------------------------------------------------===//
+include "AMDILIntrinsics.td"
+
+//===--------------------------------------------------------------------===//
+// Instructions support
+//===--------------------------------------------------------------------===//
+//===---------------------------------------------------------------------===//
+// Custom Inserter for Branches and returns, this eventually will be a
+// seperate pass
+//===---------------------------------------------------------------------===//
+let isTerminator = 1, usesCustomInserter = 1, isBranch = 1, isBarrier = 1 in {
+  def BRANCH : ILFormat<(outs), (ins brtarget:$target),
+      "; Pseudo unconditional branch instruction",
+      [(br bb:$target)]>;
+  defm BRANCH_COND : BranchConditional<IL_brcond>;
+}
+
+//===---------------------------------------------------------------------===//
+// Flow and Program control Instructions
+//===---------------------------------------------------------------------===//
+let isTerminator=1 in {
+  def SWITCH      : ILFormat< (outs), (ins GPRI32:$src),
+  !strconcat("SWITCH", " $src"), []>;
+  def CASE        : ILFormat< (outs), (ins GPRI32:$src),
+      !strconcat("CASE", " $src"), []>;
+  def BREAK       : ILFormat< (outs), (ins),
+      "BREAK", []>;
+  def CONTINUE    : ILFormat< (outs), (ins),
+      "CONTINUE", []>;
+  def DEFAULT     : ILFormat< (outs), (ins),
+      "DEFAULT", []>;
+  def ELSE        : ILFormat< (outs), (ins),
+      "ELSE", []>;
+  def ENDSWITCH   : ILFormat< (outs), (ins),
+      "ENDSWITCH", []>;
+  def ENDMAIN     : ILFormat< (outs), (ins),
+      "ENDMAIN", []>;
+  def END         : ILFormat< (outs), (ins),
+      "END", []>;
+  def ENDFUNC     : ILFormat< (outs), (ins),
+      "ENDFUNC", []>;
+  def ENDIF       : ILFormat< (outs), (ins),
+      "ENDIF", []>;
+  def WHILELOOP   : ILFormat< (outs), (ins),
+      "WHILE", []>;
+  def ENDLOOP     : ILFormat< (outs), (ins),
+      "ENDLOOP", []>;
+  def FUNC        : ILFormat< (outs), (ins),
+      "FUNC", []>;
+  def RETDYN      : ILFormat< (outs), (ins),
+      "RET_DYN", []>;
+  // This opcode has custom swizzle pattern encoded in Swizzle Encoder
+  defm IF_LOGICALNZ  : BranchInstr<"IF_LOGICALNZ">;
+  // This opcode has custom swizzle pattern encoded in Swizzle Encoder
+  defm IF_LOGICALZ   : BranchInstr<"IF_LOGICALZ">;
+  // This opcode has custom swizzle pattern encoded in Swizzle Encoder
+  defm BREAK_LOGICALNZ : BranchInstr<"BREAK_LOGICALNZ">;
+  // This opcode has custom swizzle pattern encoded in Swizzle Encoder
+  defm BREAK_LOGICALZ : BranchInstr<"BREAK_LOGICALZ">;
+  // This opcode has custom swizzle pattern encoded in Swizzle Encoder
+  defm CONTINUE_LOGICALNZ : BranchInstr<"CONTINUE_LOGICALNZ">;
+  // This opcode has custom swizzle pattern encoded in Swizzle Encoder
+  defm CONTINUE_LOGICALZ : BranchInstr<"CONTINUE_LOGICALZ">;
+  defm IFC         : BranchInstr2<"IFC">;
+  defm BREAKC      : BranchInstr2<"BREAKC">;
+  defm CONTINUEC   : BranchInstr2<"CONTINUEC">;
+}
diff --git a/lib/Target/R600/AMDILIntrinsicInfo.cpp b/lib/Target/R600/AMDILIntrinsicInfo.cpp
new file mode 100644
index 0000000000..02d06d6f3d
--- /dev/null
+++ b/lib/Target/R600/AMDILIntrinsicInfo.cpp
@@ -0,0 +1,79 @@
+//===- AMDILIntrinsicInfo.cpp - AMDGPU Intrinsic Information ------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief AMDGPU Implementation of the IntrinsicInfo class.
+//
+//===-----------------------------------------------------------------------===//
+
+#include "AMDILIntrinsicInfo.h"
+#include "AMDIL.h"
+#include "AMDGPUSubtarget.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Intrinsics.h"
+#include "llvm/Module.h"
+
+using namespace llvm;
+
+#define GET_LLVM_INTRINSIC_FOR_GCC_BUILTIN
+#include "AMDGPUGenIntrinsics.inc"
+#undef GET_LLVM_INTRINSIC_FOR_GCC_BUILTIN
+
+AMDGPUIntrinsicInfo::AMDGPUIntrinsicInfo(TargetMachine *tm) 
+  : TargetIntrinsicInfo() {
+}
+
+std::string 
+AMDGPUIntrinsicInfo::getName(unsigned int IntrID, Type **Tys,
+    unsigned int numTys) const  {
+  static const char* const names[] = {
+#define GET_INTRINSIC_NAME_TABLE
+#include "AMDGPUGenIntrinsics.inc"
+#undef GET_INTRINSIC_NAME_TABLE
+  };
+
+  if (IntrID < Intrinsic::num_intrinsics) {
+    return 0;
+  }
+  assert(IntrID < AMDGPUIntrinsic::num_AMDGPU_intrinsics
+      && "Invalid intrinsic ID");
+
+  std::string Result(names[IntrID - Intrinsic::num_intrinsics]);
+  return Result;
+}
+
+unsigned int
+AMDGPUIntrinsicInfo::lookupName(const char *Name, unsigned int Len) const  {
+#define GET_FUNCTION_RECOGNIZER
+#include "AMDGPUGenIntrinsics.inc"
+#undef GET_FUNCTION_RECOGNIZER
+  AMDGPUIntrinsic::ID IntrinsicID
+    = (AMDGPUIntrinsic::ID)Intrinsic::not_intrinsic;
+  IntrinsicID = getIntrinsicForGCCBuiltin("AMDGPU", Name);
+
+  if (IntrinsicID != (AMDGPUIntrinsic::ID)Intrinsic::not_intrinsic) {
+    return IntrinsicID;
+  }
+  return 0;
+}
+
+bool 
+AMDGPUIntrinsicInfo::isOverloaded(unsigned id) const  {
+  // Overload Table
+#define GET_INTRINSIC_OVERLOAD_TABLE
+#include "AMDGPUGenIntrinsics.inc"
+#undef GET_INTRINSIC_OVERLOAD_TABLE
+}
+
+Function*
+AMDGPUIntrinsicInfo::getDeclaration(Module *M, unsigned IntrID,
+    Type **Tys,
+    unsigned numTys) const  {
+  assert(!"Not implemented");
+}
diff --git a/lib/Target/R600/AMDILIntrinsicInfo.h b/lib/Target/R600/AMDILIntrinsicInfo.h
new file mode 100644
index 0000000000..83f4933611
--- /dev/null
+++ b/lib/Target/R600/AMDILIntrinsicInfo.h
@@ -0,0 +1,49 @@
+//===- AMDILIntrinsicInfo.h - AMDGPU Intrinsic Information ------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Interface for the AMDGPU Implementation of the Intrinsic Info class.
+//
+//===-----------------------------------------------------------------------===//
+#ifndef AMDIL_INTRINSICS_H
+#define AMDIL_INTRINSICS_H
+
+#include "llvm/Intrinsics.h"
+#include "llvm/Target/TargetIntrinsicInfo.h"
+
+namespace llvm {
+class TargetMachine;
+
+namespace AMDGPUIntrinsic {
+enum ID {
+  last_non_AMDGPU_intrinsic = Intrinsic::num_intrinsics - 1,
+#define GET_INTRINSIC_ENUM_VALUES
+#include "AMDGPUGenIntrinsics.inc"
+#undef GET_INTRINSIC_ENUM_VALUES
+      , num_AMDGPU_intrinsics
+};
+
+} // end namespace AMDGPUIntrinsic
+
+class AMDGPUIntrinsicInfo : public TargetIntrinsicInfo {
+public:
+  AMDGPUIntrinsicInfo(TargetMachine *tm);
+  std::string getName(unsigned int IntrId, Type **Tys = 0,
+                      unsigned int numTys = 0) const;
+  unsigned int lookupName(const char *Name, unsigned int Len) const;
+  bool isOverloaded(unsigned int IID) const;
+  Function *getDeclaration(Module *M, unsigned int ID,
+                           Type **Tys = 0,
+                           unsigned int numTys = 0) const;
+};
+
+} // end namespace llvm
+
+#endif // AMDIL_INTRINSICS_H
+
diff --git a/lib/Target/R600/AMDILIntrinsics.td b/lib/Target/R600/AMDILIntrinsics.td
new file mode 100644
index 0000000000..3f9e20f0c8
--- /dev/null
+++ b/lib/Target/R600/AMDILIntrinsics.td
@@ -0,0 +1,242 @@
+//===- AMDILIntrinsics.td - Defines AMDIL Intrinscs -*- tablegen -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+// This file defines all of the amdil-specific intrinsics
+//
+//===---------------------------------------------------------------===//
+//===--------------------------------------------------------------------===//
+// Intrinsic classes
+// Generic versions of the above classes but for Target specific intrinsics
+// instead of SDNode patterns.
+//===--------------------------------------------------------------------===//
+let TargetPrefix = "AMDIL", isTarget = 1 in {
+     class VoidIntLong :
+          Intrinsic<[llvm_i64_ty], [], []>;
+     class VoidIntInt :
+          Intrinsic<[llvm_i32_ty], [], []>;
+     class VoidIntBool :
+          Intrinsic<[llvm_i32_ty], [], []>;
+     class UnaryIntInt :
+          Intrinsic<[llvm_anyint_ty], [LLVMMatchType<0>], [IntrNoMem]>;
+     class UnaryIntFloat :
+          Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
+     class ConvertIntFTOI :
+          Intrinsic<[llvm_anyint_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
+     class ConvertIntITOF :
+          Intrinsic<[llvm_anyfloat_ty], [llvm_anyint_ty], [IntrNoMem]>;
+     class UnaryIntNoRetInt :
+          Intrinsic<[], [llvm_anyint_ty], []>;
+     class UnaryIntNoRetFloat :
+          Intrinsic<[], [llvm_anyfloat_ty], []>;
+     class BinaryIntInt :
+          Intrinsic<[llvm_anyint_ty], [LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
+     class BinaryIntFloat :
+          Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
+     class BinaryIntNoRetInt :
+          Intrinsic<[], [llvm_anyint_ty, LLVMMatchType<0>], []>;
+     class BinaryIntNoRetFloat :
+          Intrinsic<[], [llvm_anyfloat_ty, LLVMMatchType<0>], []>;
+     class TernaryIntInt :
+          Intrinsic<[llvm_anyint_ty], [LLVMMatchType<0>,
+          LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
+     class TernaryIntFloat :
+          Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>,
+          LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
+     class QuaternaryIntInt :
+          Intrinsic<[llvm_anyint_ty], [LLVMMatchType<0>,
+          LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>;
+     class UnaryAtomicInt :
+          Intrinsic<[llvm_i32_ty], [llvm_ptr_ty, llvm_i32_ty], [IntrReadWriteArgMem]>;
+     class BinaryAtomicInt :
+          Intrinsic<[llvm_i32_ty], [llvm_ptr_ty, llvm_i32_ty, llvm_i32_ty], [IntrReadWriteArgMem]>;
+     class TernaryAtomicInt :
+          Intrinsic<[llvm_i32_ty], [llvm_ptr_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty]>;
+     class UnaryAtomicIntNoRet :
+          Intrinsic<[], [llvm_ptr_ty, llvm_i32_ty], [IntrReadWriteArgMem]>;
+     class BinaryAtomicIntNoRet :
+          Intrinsic<[], [llvm_ptr_ty, llvm_i32_ty, llvm_i32_ty], [IntrReadWriteArgMem]>;
+     class TernaryAtomicIntNoRet :
+          Intrinsic<[], [llvm_ptr_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrReadWriteArgMem]>;
+}
+
+let TargetPrefix = "AMDIL", isTarget = 1 in {
+  def int_AMDIL_abs : GCCBuiltin<"__amdil_abs">, UnaryIntInt;
+
+  def int_AMDIL_bit_extract_i32 : GCCBuiltin<"__amdil_ibit_extract">,
+          TernaryIntInt;
+  def int_AMDIL_bit_extract_u32 : GCCBuiltin<"__amdil_ubit_extract">,
+          TernaryIntInt;
+  def int_AMDIL_bit_reverse_u32 : GCCBuiltin<"__amdil_ubit_reverse">,
+          UnaryIntInt;
+  def int_AMDIL_bit_count_i32 : GCCBuiltin<"__amdil_count_bits">,
+          UnaryIntInt;
+  def int_AMDIL_bit_find_first_lo : GCCBuiltin<"__amdil_ffb_lo">,
+          UnaryIntInt;
+  def int_AMDIL_bit_find_first_hi : GCCBuiltin<"__amdil_ffb_hi">,
+          UnaryIntInt;
+  def int_AMDIL_bit_find_first_sgn : GCCBuiltin<"__amdil_ffb_signed">,
+          UnaryIntInt;
+  def int_AMDIL_media_bitalign : GCCBuiltin<"__amdil_bitalign">,
+                    TernaryIntInt;
+  def int_AMDIL_media_bytealign : GCCBuiltin<"__amdil_bytealign">,
+                    TernaryIntInt;
+  def int_AMDIL_bit_insert_u32 : GCCBuiltin<"__amdil_ubit_insert">,
+                    QuaternaryIntInt;
+  def int_AMDIL_bfi : GCCBuiltin<"__amdil_bfi">,
+      TernaryIntInt;
+  def int_AMDIL_bfm : GCCBuiltin<"__amdil_bfm">,
+      BinaryIntInt;
+  def int_AMDIL_mad_i32 : GCCBuiltin<"__amdil_imad">,
+          TernaryIntInt;
+  def int_AMDIL_mad_u32 : GCCBuiltin<"__amdil_umad">,
+          TernaryIntInt;
+  def int_AMDIL_mad     : GCCBuiltin<"__amdil_mad">,
+          TernaryIntFloat;
+  def int_AMDIL_mulhi_i32 : GCCBuiltin<"__amdil_imul_high">,
+          BinaryIntInt;
+  def int_AMDIL_mulhi_u32 : GCCBuiltin<"__amdil_umul_high">,
+          BinaryIntInt;
+  def int_AMDIL_mul24_i32 : GCCBuiltin<"__amdil_imul24">,
+          BinaryIntInt;
+  def int_AMDIL_mul24_u32 : GCCBuiltin<"__amdil_umul24">,
+          BinaryIntInt;
+  def int_AMDIL_mulhi24_i32 : GCCBuiltin<"__amdil_imul24_high">,
+          BinaryIntInt;
+  def int_AMDIL_mulhi24_u32 : GCCBuiltin<"__amdil_umul24_high">,
+          BinaryIntInt;
+  def int_AMDIL_mad24_i32 : GCCBuiltin<"__amdil_imad24">,
+          TernaryIntInt;
+  def int_AMDIL_mad24_u32 : GCCBuiltin<"__amdil_umad24">,
+          TernaryIntInt;
+  def int_AMDIL_carry_i32 : GCCBuiltin<"__amdil_carry">,
+          BinaryIntInt;
+  def int_AMDIL_borrow_i32 : GCCBuiltin<"__amdil_borrow">,
+          BinaryIntInt;
+  def int_AMDIL_min_i32 : GCCBuiltin<"__amdil_imin">,
+          BinaryIntInt;
+  def int_AMDIL_min_u32 : GCCBuiltin<"__amdil_umin">,
+          BinaryIntInt;
+  def int_AMDIL_min     : GCCBuiltin<"__amdil_min">,
+          BinaryIntFloat;
+  def int_AMDIL_max_i32 : GCCBuiltin<"__amdil_imax">,
+          BinaryIntInt;
+  def int_AMDIL_max_u32 : GCCBuiltin<"__amdil_umax">,
+          BinaryIntInt;
+  def int_AMDIL_max     : GCCBuiltin<"__amdil_max">,
+          BinaryIntFloat;
+  def int_AMDIL_media_lerp_u4 : GCCBuiltin<"__amdil_u4lerp">,
+          TernaryIntInt;
+  def int_AMDIL_media_sad : GCCBuiltin<"__amdil_sad">,
+          TernaryIntInt;
+  def int_AMDIL_media_sad_hi : GCCBuiltin<"__amdil_sadhi">,
+          TernaryIntInt;
+  def int_AMDIL_fraction : GCCBuiltin<"__amdil_fraction">,
+          UnaryIntFloat;
+  def int_AMDIL_clamp : GCCBuiltin<"__amdil_clamp">,
+          TernaryIntFloat;
+  def int_AMDIL_pireduce : GCCBuiltin<"__amdil_pireduce">,
+          UnaryIntFloat;
+  def int_AMDIL_round_nearest : GCCBuiltin<"__amdil_round_nearest">,
+          UnaryIntFloat;
+  def int_AMDIL_round_neginf : GCCBuiltin<"__amdil_round_neginf">,
+          UnaryIntFloat;
+  def int_AMDIL_round_zero : GCCBuiltin<"__amdil_round_zero">,
+          UnaryIntFloat;
+  def int_AMDIL_acos : GCCBuiltin<"__amdil_acos">,
+          UnaryIntFloat;
+  def int_AMDIL_atan : GCCBuiltin<"__amdil_atan">,
+          UnaryIntFloat;
+  def int_AMDIL_asin : GCCBuiltin<"__amdil_asin">,
+          UnaryIntFloat;
+  def int_AMDIL_cos : GCCBuiltin<"__amdil_cos">,
+          UnaryIntFloat;
+  def int_AMDIL_cos_vec : GCCBuiltin<"__amdil_cos_vec">,
+          UnaryIntFloat;
+  def int_AMDIL_tan : GCCBuiltin<"__amdil_tan">,
+          UnaryIntFloat;
+  def int_AMDIL_sin : GCCBuiltin<"__amdil_sin">,
+          UnaryIntFloat;
+  def int_AMDIL_sin_vec : GCCBuiltin<"__amdil_sin_vec">,
+          UnaryIntFloat;
+  def int_AMDIL_pow : GCCBuiltin<"__amdil_pow">, BinaryIntFloat;
+  def int_AMDIL_div : GCCBuiltin<"__amdil_div">, BinaryIntFloat;
+  def int_AMDIL_udiv : GCCBuiltin<"__amdil_udiv">, BinaryIntInt;
+  def int_AMDIL_sqrt: GCCBuiltin<"__amdil_sqrt">,
+          UnaryIntFloat;
+  def int_AMDIL_sqrt_vec: GCCBuiltin<"__amdil_sqrt_vec">,
+          UnaryIntFloat;
+  def int_AMDIL_exp : GCCBuiltin<"__amdil_exp">,
+          UnaryIntFloat;
+  def int_AMDIL_exp_vec : GCCBuiltin<"__amdil_exp_vec">,
+          UnaryIntFloat;
+  def int_AMDIL_exn : GCCBuiltin<"__amdil_exn">,
+          UnaryIntFloat;
+  def int_AMDIL_log_vec : GCCBuiltin<"__amdil_log_vec">,
+          UnaryIntFloat;
+  def int_AMDIL_ln : GCCBuiltin<"__amdil_ln">,
+          UnaryIntFloat;
+  def int_AMDIL_sign: GCCBuiltin<"__amdil_sign">,
+          UnaryIntFloat;
+  def int_AMDIL_fma: GCCBuiltin<"__amdil_fma">,
+          TernaryIntFloat;
+  def int_AMDIL_rsq : GCCBuiltin<"__amdil_rsq">,
+          UnaryIntFloat;
+  def int_AMDIL_rsq_vec : GCCBuiltin<"__amdil_rsq_vec">,
+          UnaryIntFloat;
+  def int_AMDIL_length : GCCBuiltin<"__amdil_length">,
+          UnaryIntFloat;
+  def int_AMDIL_lerp : GCCBuiltin<"__amdil_lerp">,
+          TernaryIntFloat;
+  def int_AMDIL_media_sad4 : GCCBuiltin<"__amdil_sad4">,
+      Intrinsic<[llvm_i32_ty], [llvm_v4i32_ty,
+           llvm_v4i32_ty, llvm_i32_ty], []>;
+
+  def int_AMDIL_frexp_f64 : GCCBuiltin<"__amdil_frexp">,
+        Intrinsic<[llvm_v2i64_ty], [llvm_double_ty], []>;
+ def int_AMDIL_ldexp : GCCBuiltin<"__amdil_ldexp">,
+    Intrinsic<[llvm_anyfloat_ty], [llvm_anyfloat_ty, llvm_anyint_ty], []>;
+  def int_AMDIL_drcp : GCCBuiltin<"__amdil_rcp">,
+      Intrinsic<[llvm_double_ty], [llvm_double_ty], []>;
+  def int_AMDIL_convert_f16_f32 : GCCBuiltin<"__amdil_half_to_float">,
+      ConvertIntITOF;
+  def int_AMDIL_convert_f32_f16 : GCCBuiltin<"__amdil_float_to_half">,
+      ConvertIntFTOI;
+  def int_AMDIL_convert_f32_i32_rpi : GCCBuiltin<"__amdil_float_to_int_rpi">,
+      ConvertIntFTOI;
+  def int_AMDIL_convert_f32_i32_flr : GCCBuiltin<"__amdil_float_to_int_flr">,
+      ConvertIntFTOI;
+  def int_AMDIL_convert_f32_f16_near : GCCBuiltin<"__amdil_float_to_half_near">,
+      ConvertIntFTOI;
+  def int_AMDIL_convert_f32_f16_neg_inf : GCCBuiltin<"__amdil_float_to_half_neg_inf">,
+      ConvertIntFTOI;
+  def int_AMDIL_convert_f32_f16_plus_inf : GCCBuiltin<"__amdil_float_to_half_plus_inf">,
+      ConvertIntFTOI;
+ def int_AMDIL_media_convert_f2v4u8 : GCCBuiltin<"__amdil_f_2_u4">,
+      Intrinsic<[llvm_i32_ty], [llvm_v4f32_ty], []>;
+  def int_AMDIL_media_unpack_byte_0 : GCCBuiltin<"__amdil_unpack_0">,
+      ConvertIntITOF;
+  def int_AMDIL_media_unpack_byte_1 : GCCBuiltin<"__amdil_unpack_1">,
+      ConvertIntITOF;
+  def int_AMDIL_media_unpack_byte_2 : GCCBuiltin<"__amdil_unpack_2">,
+      ConvertIntITOF;
+  def int_AMDIL_media_unpack_byte_3 : GCCBuiltin<"__amdil_unpack_3">,
+      ConvertIntITOF;
+  def int_AMDIL_dp2_add : GCCBuiltin<"__amdil_dp2_add">,
+        Intrinsic<[llvm_float_ty], [llvm_v2f32_ty,
+          llvm_v2f32_ty, llvm_float_ty], []>;
+  def int_AMDIL_dp2 : GCCBuiltin<"__amdil_dp2">,
+        Intrinsic<[llvm_float_ty], [llvm_v2f32_ty,
+          llvm_v2f32_ty], []>;
+  def int_AMDIL_dp3 : GCCBuiltin<"__amdil_dp3">,
+        Intrinsic<[llvm_float_ty], [llvm_v4f32_ty,
+          llvm_v4f32_ty], []>;
+  def int_AMDIL_dp4 : GCCBuiltin<"__amdil_dp4">,
+        Intrinsic<[llvm_float_ty], [llvm_v4f32_ty,
+          llvm_v4f32_ty], []>;
+}
diff --git a/lib/Target/R600/AMDILNIDevice.cpp b/lib/Target/R600/AMDILNIDevice.cpp
new file mode 100644
index 0000000000..b82da5908a
--- /dev/null
+++ b/lib/Target/R600/AMDILNIDevice.cpp
@@ -0,0 +1,65 @@
+//===-- AMDILNIDevice.cpp - Device Info for Northern Islands devices ------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//==-----------------------------------------------------------------------===//
+#include "AMDILNIDevice.h"
+#include "AMDILEvergreenDevice.h"
+#include "AMDGPUSubtarget.h"
+
+using namespace llvm;
+
+AMDGPUNIDevice::AMDGPUNIDevice(AMDGPUSubtarget *ST)
+  : AMDGPUEvergreenDevice(ST) {
+  std::string name = ST->getDeviceName();
+  if (name == "caicos") {
+    DeviceFlag = OCL_DEVICE_CAICOS;
+  } else if (name == "turks") {
+    DeviceFlag = OCL_DEVICE_TURKS;
+  } else if (name == "cayman") {
+    DeviceFlag = OCL_DEVICE_CAYMAN;
+  } else {
+    DeviceFlag = OCL_DEVICE_BARTS;
+  }
+}
+AMDGPUNIDevice::~AMDGPUNIDevice() {
+}
+
+size_t
+AMDGPUNIDevice::getMaxLDSSize() const {
+  if (usesHardware(AMDGPUDeviceInfo::LocalMem)) {
+    return MAX_LDS_SIZE_900;
+  } else {
+    return 0;
+  }
+}
+
+uint32_t
+AMDGPUNIDevice::getGeneration() const {
+  return AMDGPUDeviceInfo::HD6XXX;
+}
+
+
+AMDGPUCaymanDevice::AMDGPUCaymanDevice(AMDGPUSubtarget *ST)
+  : AMDGPUNIDevice(ST) {
+  setCaps();
+}
+
+AMDGPUCaymanDevice::~AMDGPUCaymanDevice() {
+}
+
+void
+AMDGPUCaymanDevice::setCaps() {
+  if (mSTM->isOverride(AMDGPUDeviceInfo::DoubleOps)) {
+    mHWBits.set(AMDGPUDeviceInfo::DoubleOps);
+    mHWBits.set(AMDGPUDeviceInfo::FMA);
+  }
+  mHWBits.set(AMDGPUDeviceInfo::Signed24BitOps);
+  mSWBits.reset(AMDGPUDeviceInfo::Signed24BitOps);
+  mSWBits.set(AMDGPUDeviceInfo::ArenaSegment);
+}
+
diff --git a/lib/Target/R600/AMDILNIDevice.h b/lib/Target/R600/AMDILNIDevice.h
new file mode 100644
index 0000000000..bc7df37b26
--- /dev/null
+++ b/lib/Target/R600/AMDILNIDevice.h
@@ -0,0 +1,57 @@
+//===------- AMDILNIDevice.h - Define NI Device for AMDIL -*- C++ -*------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+/// \file
+/// \brief Interface for the subtarget data classes.
+///
+/// This file will define the interface that each generation needs to
+/// implement in order to correctly answer queries on the capabilities of the
+/// specific hardware.
+//===---------------------------------------------------------------------===//
+#ifndef AMDILNIDEVICE_H
+#define AMDILNIDEVICE_H
+#include "AMDILEvergreenDevice.h"
+#include "AMDGPUSubtarget.h"
+
+namespace llvm {
+
+class AMDGPUSubtarget;
+//===---------------------------------------------------------------------===//
+// NI generation of devices and their respective sub classes
+//===---------------------------------------------------------------------===//
+
+/// \brief The AMDGPUNIDevice is the base class for all Northern Island series of
+/// cards.
+///
+/// It is very similiar to the AMDGPUEvergreenDevice, with the major
+/// exception being differences in wavefront size and hardware capabilities.  The
+/// NI devices are all 64 wide wavefronts and also add support for signed 24 bit
+/// integer operations
+class AMDGPUNIDevice : public AMDGPUEvergreenDevice {
+public:
+  AMDGPUNIDevice(AMDGPUSubtarget*);
+  virtual ~AMDGPUNIDevice();
+  virtual size_t getMaxLDSSize() const;
+  virtual uint32_t getGeneration() const;
+};
+
+/// Just as the AMDGPUCypressDevice is the double capable version of the
+/// AMDGPUEvergreenDevice, the AMDGPUCaymanDevice is the double capable version
+/// of the AMDGPUNIDevice.  The other major difference is that the Cayman Device
+/// has 4 wide ALU's, whereas the rest of the NI family is a 5 wide.
+class AMDGPUCaymanDevice: public AMDGPUNIDevice {
+public:
+  AMDGPUCaymanDevice(AMDGPUSubtarget*);
+  virtual ~AMDGPUCaymanDevice();
+private:
+  virtual void setCaps();
+};
+
+static const unsigned int MAX_LDS_SIZE_900 = AMDGPUDevice::MAX_LDS_SIZE_800;
+} // namespace llvm
+#endif // AMDILNIDEVICE_H
diff --git a/lib/Target/R600/AMDILPeepholeOptimizer.cpp b/lib/Target/R600/AMDILPeepholeOptimizer.cpp
new file mode 100644
index 0000000000..4a748b8e9c
--- /dev/null
+++ b/lib/Target/R600/AMDILPeepholeOptimizer.cpp
@@ -0,0 +1,1215 @@
+//===-- AMDILPeepholeOptimizer.cpp - AMDGPU Peephole optimizations ---------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//==-----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "PeepholeOpt"
+#ifdef DEBUG
+#define DEBUGME (DebugFlag && isCurrentDebugType(DEBUG_TYPE))
+#else
+#define DEBUGME 0
+#endif
+
+#include "AMDILDevices.h"
+#include "AMDGPUInstrInfo.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/StringExtras.h"
+#include "llvm/ADT/StringRef.h"
+#include "llvm/ADT/Twine.h"
+#include "llvm/Constants.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFunctionAnalysis.h"
+#include "llvm/Function.h"
+#include "llvm/Instructions.h"
+#include "llvm/Module.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/MathExtras.h"
+
+#include <sstream>
+
+#if 0
+STATISTIC(PointerAssignments, "Number of dynamic pointer "
+    "assigments discovered");
+STATISTIC(PointerSubtract, "Number of pointer subtractions discovered");
+#endif
+
+using namespace llvm;
+// The Peephole optimization pass is used to do simple last minute optimizations
+// that are required for correct code or to remove redundant functions
+namespace {
+
+class OpaqueType;
+
+class LLVM_LIBRARY_VISIBILITY AMDGPUPeepholeOpt : public FunctionPass {
+public:
+  TargetMachine &TM;
+  static char ID;
+  AMDGPUPeepholeOpt(TargetMachine &tm);
+  ~AMDGPUPeepholeOpt();
+  const char *getPassName() const;
+  bool runOnFunction(Function &F);
+  bool doInitialization(Module &M);
+  bool doFinalization(Module &M);
+  void getAnalysisUsage(AnalysisUsage &AU) const;
+protected:
+private:
+  // Function to initiate all of the instruction level optimizations.
+  bool instLevelOptimizations(BasicBlock::iterator *inst);
+  // Quick check to see if we need to dump all of the pointers into the
+  // arena. If this is correct, then we set all pointers to exist in arena. This
+  // is a workaround for aliasing of pointers in a struct/union.
+  bool dumpAllIntoArena(Function &F);
+  // Because I don't want to invalidate any pointers while in the
+  // safeNestedForEachFunction. I push atomic conversions to a vector and handle
+  // it later. This function does the conversions if required.
+  void doAtomicConversionIfNeeded(Function &F);
+  // Because __amdil_is_constant cannot be properly evaluated if
+  // optimizations are disabled, the call's are placed in a vector
+  // and evaluated after the __amdil_image* functions are evaluated
+  // which should allow the __amdil_is_constant function to be
+  // evaluated correctly.
+  void doIsConstCallConversionIfNeeded();
+  bool mChanged;
+  bool mDebug;
+  bool mConvertAtomics;
+  CodeGenOpt::Level optLevel;
+  // Run a series of tests to see if we can optimize a CALL instruction.
+  bool optimizeCallInst(BasicBlock::iterator *bbb);
+  // A peephole optimization to optimize bit extract sequences.
+  bool optimizeBitExtract(Instruction *inst);
+  // A peephole optimization to optimize bit insert sequences.
+  bool optimizeBitInsert(Instruction *inst);
+  bool setupBitInsert(Instruction *base, 
+                      Instruction *&src, 
+                      Constant *&mask, 
+                      Constant *&shift);
+  // Expand the bit field insert instruction on versions of OpenCL that
+  // don't support it.
+  bool expandBFI(CallInst *CI);
+  // Expand the bit field mask instruction on version of OpenCL that 
+  // don't support it.
+  bool expandBFM(CallInst *CI);
+  // On 7XX and 8XX operations, we do not have 24 bit signed operations. So in
+  // this case we need to expand them. These functions check for 24bit functions
+  // and then expand.
+  bool isSigned24BitOps(CallInst *CI);
+  void expandSigned24BitOps(CallInst *CI);
+  // One optimization that can occur is that if the required workgroup size is
+  // specified then the result of get_local_size is known at compile time and
+  // can be returned accordingly.
+  bool isRWGLocalOpt(CallInst *CI);
+  // On northern island cards, the division is slightly less accurate than on
+  // previous generations, so we need to utilize a more accurate division. So we
+  // can translate the accurate divide to a normal divide on all other cards.
+  bool convertAccurateDivide(CallInst *CI);
+  void expandAccurateDivide(CallInst *CI);
+  // If the alignment is set incorrectly, it can produce really inefficient
+  // code. This checks for this scenario and fixes it if possible.
+  bool correctMisalignedMemOp(Instruction *inst);
+
+  // If we are in no opt mode, then we need to make sure that
+  // local samplers are properly propagated as constant propagation 
+  // doesn't occur and we need to know the value of kernel defined
+  // samplers at compile time.
+  bool propagateSamplerInst(CallInst *CI);
+
+  // Helper functions
+
+  // Group of functions that recursively calculate the size of a structure based
+  // on it's sub-types.
+  size_t getTypeSize(Type * const T, bool dereferencePtr = false);
+  size_t getTypeSize(StructType * const ST, bool dereferencePtr = false);
+  size_t getTypeSize(IntegerType * const IT, bool dereferencePtr = false);
+  size_t getTypeSize(FunctionType * const FT,bool dereferencePtr = false);
+  size_t getTypeSize(ArrayType * const AT, bool dereferencePtr = false);
+  size_t getTypeSize(VectorType * const VT, bool dereferencePtr = false);
+  size_t getTypeSize(PointerType * const PT, bool dereferencePtr = false);
+  size_t getTypeSize(OpaqueType * const OT, bool dereferencePtr = false);
+
+  LLVMContext *mCTX;
+  Function *mF;
+  const AMDGPUSubtarget *mSTM;
+  SmallVector< std::pair<CallInst *, Function *>, 16> atomicFuncs;
+  SmallVector<CallInst *, 16> isConstVec;
+}; // class AMDGPUPeepholeOpt
+  char AMDGPUPeepholeOpt::ID = 0;
+
+// A template function that has two levels of looping before calling the
+// function with a pointer to the current iterator.
+template<class InputIterator, class SecondIterator, class Function>
+Function safeNestedForEach(InputIterator First, InputIterator Last,
+                              SecondIterator S, Function F) {
+  for ( ; First != Last; ++First) {
+    SecondIterator sf, sl;
+    for (sf = First->begin(), sl = First->end();
+         sf != sl; )  {
+      if (!F(&sf)) {
+        ++sf;
+      } 
+    }
+  }
+  return F;
+}
+
+} // anonymous namespace
+
+namespace llvm {
+  FunctionPass *
+  createAMDGPUPeepholeOpt(TargetMachine &tm) {
+    return new AMDGPUPeepholeOpt(tm);
+  }
+} // llvm namespace
+
+AMDGPUPeepholeOpt::AMDGPUPeepholeOpt(TargetMachine &tm)
+  : FunctionPass(ID), TM(tm)  {
+  mDebug = DEBUGME;
+  optLevel = TM.getOptLevel();
+
+}
+
+AMDGPUPeepholeOpt::~AMDGPUPeepholeOpt()  {
+}
+
+const char *
+AMDGPUPeepholeOpt::getPassName() const  {
+  return "AMDGPU PeepHole Optimization Pass";
+}
+
+bool 
+containsPointerType(Type *Ty)  {
+  if (!Ty) {
+    return false;
+  }
+  switch(Ty->getTypeID()) {
+  default:
+    return false;
+  case Type::StructTyID: {
+    const StructType *ST = dyn_cast<StructType>(Ty);
+    for (StructType::element_iterator stb = ST->element_begin(),
+           ste = ST->element_end(); stb != ste; ++stb) {
+      if (!containsPointerType(*stb)) {
+        continue;
+      }
+      return true;
+    }
+    break;
+  }
+  case Type::VectorTyID:
+  case Type::ArrayTyID:
+    return containsPointerType(dyn_cast<SequentialType>(Ty)->getElementType());
+  case Type::PointerTyID:
+    return true;
+  };
+  return false;
+}
+
+bool 
+AMDGPUPeepholeOpt::dumpAllIntoArena(Function &F)  {
+  bool dumpAll = false;
+  for (Function::const_arg_iterator cab = F.arg_begin(),
+       cae = F.arg_end(); cab != cae; ++cab) {
+    const Argument *arg = cab;
+    const PointerType *PT = dyn_cast<PointerType>(arg->getType());
+    if (!PT) {
+      continue;
+    }
+    Type *DereferencedType = PT->getElementType();
+    if (!dyn_cast<StructType>(DereferencedType) 
+        ) {
+      continue;
+    }
+    if (!containsPointerType(DereferencedType)) {
+      continue;
+    }
+    // FIXME: Because a pointer inside of a struct/union may be aliased to
+    // another pointer we need to take the conservative approach and place all
+    // pointers into the arena until more advanced detection is implemented.
+    dumpAll = true;
+  }
+  return dumpAll;
+}
+void
+AMDGPUPeepholeOpt::doIsConstCallConversionIfNeeded() {
+  if (isConstVec.empty()) {
+    return;
+  }
+  for (unsigned x = 0, y = isConstVec.size(); x < y; ++x) {
+    CallInst *CI = isConstVec[x];
+    Constant *CV = dyn_cast<Constant>(CI->getOperand(0));
+    Type *aType = Type::getInt32Ty(*mCTX);
+    Value *Val = (CV != NULL) ? ConstantInt::get(aType, 1)
+      : ConstantInt::get(aType, 0);
+    CI->replaceAllUsesWith(Val);
+    CI->eraseFromParent();
+  }
+  isConstVec.clear();
+}
+void 
+AMDGPUPeepholeOpt::doAtomicConversionIfNeeded(Function &F)  {
+  // Don't do anything if we don't have any atomic operations.
+  if (atomicFuncs.empty()) {
+    return;
+  }
+  // Change the function name for the atomic if it is required
+  uint32_t size = atomicFuncs.size();
+  for (uint32_t x = 0; x < size; ++x) {
+    atomicFuncs[x].first->setOperand(
+        atomicFuncs[x].first->getNumOperands()-1, 
+        atomicFuncs[x].second);
+
+  }
+  mChanged = true;
+  if (mConvertAtomics) {
+    return;
+  }
+}
+
+bool 
+AMDGPUPeepholeOpt::runOnFunction(Function &MF)  {
+  mChanged = false;
+  mF = &MF;
+  mSTM = &TM.getSubtarget<AMDGPUSubtarget>();
+  if (mDebug) {
+    MF.dump();
+  }
+  mCTX = &MF.getType()->getContext();
+  mConvertAtomics = true;
+  safeNestedForEach(MF.begin(), MF.end(), MF.begin()->begin(),
+     std::bind1st(std::mem_fun(&AMDGPUPeepholeOpt::instLevelOptimizations),
+                  this));
+
+  doAtomicConversionIfNeeded(MF);
+  doIsConstCallConversionIfNeeded();
+
+  if (mDebug) {
+    MF.dump();
+  }
+  return mChanged;
+}
+
+bool 
+AMDGPUPeepholeOpt::optimizeCallInst(BasicBlock::iterator *bbb)  {
+  Instruction *inst = (*bbb);
+  CallInst *CI = dyn_cast<CallInst>(inst);
+  if (!CI) {
+    return false;
+  }
+  if (isSigned24BitOps(CI)) {
+    expandSigned24BitOps(CI);
+    ++(*bbb);
+    CI->eraseFromParent();
+    return true;
+  }
+  if (propagateSamplerInst(CI)) {
+    return false;
+  }
+  if (expandBFI(CI) || expandBFM(CI)) {
+    ++(*bbb);
+    CI->eraseFromParent();
+    return true;
+  }
+  if (convertAccurateDivide(CI)) {
+    expandAccurateDivide(CI);
+    ++(*bbb);
+    CI->eraseFromParent();
+    return true;
+  }
+
+  StringRef calleeName = CI->getOperand(CI->getNumOperands()-1)->getName();
+  if (calleeName.startswith("__amdil_is_constant")) {
+    // If we do not have optimizations, then this
+    // cannot be properly evaluated, so we add the
+    // call instruction to a vector and process
+    // them at the end of processing after the
+    // samplers have been correctly handled.
+    if (optLevel == CodeGenOpt::None) {
+      isConstVec.push_back(CI);
+      return false;
+    } else {
+      Constant *CV = dyn_cast<Constant>(CI->getOperand(0));
+      Type *aType = Type::getInt32Ty(*mCTX);
+      Value *Val = (CV != NULL) ? ConstantInt::get(aType, 1)
+        : ConstantInt::get(aType, 0);
+      CI->replaceAllUsesWith(Val);
+      ++(*bbb);
+      CI->eraseFromParent();
+      return true;
+    }
+  }
+
+  if (calleeName.equals("__amdil_is_asic_id_i32")) {
+    ConstantInt *CV = dyn_cast<ConstantInt>(CI->getOperand(0));
+    Type *aType = Type::getInt32Ty(*mCTX);
+    Value *Val = CV;
+    if (Val) {
+      Val = ConstantInt::get(aType, 
+          mSTM->device()->getDeviceFlag() & CV->getZExtValue());
+    } else {
+      Val = ConstantInt::get(aType, 0);
+    }
+    CI->replaceAllUsesWith(Val);
+    ++(*bbb);
+    CI->eraseFromParent();
+    return true;
+  }
+  Function *F = dyn_cast<Function>(CI->getOperand(CI->getNumOperands()-1));
+  if (!F) {
+    return false;
+  } 
+  if (F->getName().startswith("__atom") && !CI->getNumUses() 
+      && F->getName().find("_xchg") == StringRef::npos) {
+    std::string buffer(F->getName().str() + "_noret");
+    F = dyn_cast<Function>(
+          F->getParent()->getOrInsertFunction(buffer, F->getFunctionType()));
+    atomicFuncs.push_back(std::make_pair <CallInst*, Function*>(CI, F));
+  }
+  
+  if (!mSTM->device()->isSupported(AMDGPUDeviceInfo::ArenaSegment)
+      && !mSTM->device()->isSupported(AMDGPUDeviceInfo::MultiUAV)) {
+    return false;
+  }
+  if (!mConvertAtomics) {
+    return false;
+  }
+  StringRef name = F->getName();
+  if (name.startswith("__atom") && name.find("_g") != StringRef::npos) {
+    mConvertAtomics = false;
+  }
+  return false;
+}
+
+bool
+AMDGPUPeepholeOpt::setupBitInsert(Instruction *base, 
+    Instruction *&src, 
+    Constant *&mask, 
+    Constant *&shift) {
+  if (!base) {
+    if (mDebug) {
+      dbgs() << "Null pointer passed into function.\n";
+    }
+    return false;
+  }
+  bool andOp = false;
+  if (base->getOpcode() == Instruction::Shl) {
+    shift = dyn_cast<Constant>(base->getOperand(1));
+  } else if (base->getOpcode() == Instruction::And) {
+    mask = dyn_cast<Constant>(base->getOperand(1));
+    andOp = true;
+  } else {
+    if (mDebug) {
+      dbgs() << "Failed setup with no Shl or And instruction on base opcode!\n";
+    }
+    // If the base is neither a Shl or a And, we don't fit any of the patterns above.
+    return false;
+  }
+  src = dyn_cast<Instruction>(base->getOperand(0));
+  if (!src) {
+    if (mDebug) {
+      dbgs() << "Failed setup since the base operand is not an instruction!\n";
+    }
+    return false;
+  }
+  // If we find an 'and' operation, then we don't need to
+  // find the next operation as we already know the
+  // bits that are valid at this point.
+  if (andOp) {
+    return true;
+  }
+  if (src->getOpcode() == Instruction::Shl && !shift) {
+    shift = dyn_cast<Constant>(src->getOperand(1));
+    src = dyn_cast<Instruction>(src->getOperand(0));
+  } else if (src->getOpcode() == Instruction::And && !mask) {
+    mask = dyn_cast<Constant>(src->getOperand(1));
+  }
+  if (!mask && !shift) {
+    if (mDebug) {
+      dbgs() << "Failed setup since both mask and shift are NULL!\n";
+    }
+    // Did not find a constant mask or a shift.
+    return false;
+  }
+  return true;
+}
+bool
+AMDGPUPeepholeOpt::optimizeBitInsert(Instruction *inst)  {
+  if (!inst) {
+    return false;
+  }
+  if (!inst->isBinaryOp()) {
+    return false;
+  }
+  if (inst->getOpcode() != Instruction::Or) {
+    return false;
+  }
+  if (optLevel == CodeGenOpt::None) {
+    return false;
+  }
+  // We want to do an optimization on a sequence of ops that in the end equals a
+  // single ISA instruction.
+  // The base pattern for this optimization is - ((A & B) << C) | ((D & E) << F)
+  // Some simplified versions of this pattern are as follows:
+  // (A & B) | (D & E) when B & E == 0 && C == 0 && F == 0
+  // ((A & B) << C) | (D & E) when B ^ E == 0 && (1 << C) >= E
+  // (A & B) | ((D & E) << F) when B ^ E == 0 && (1 << F) >= B
+  // (A & B) | (D << F) when (1 << F) >= B
+  // (A << C) | (D & E) when (1 << C) >= E
+  if (mSTM->device()->getGeneration() == AMDGPUDeviceInfo::HD4XXX) {
+    // The HD4XXX hardware doesn't support the ubit_insert instruction.
+    return false;
+  }
+  Type *aType = inst->getType();
+  bool isVector = aType->isVectorTy();
+  int numEle = 1;
+  // This optimization only works on 32bit integers.
+  if (aType->getScalarType()
+      != Type::getInt32Ty(inst->getContext())) {
+    return false;
+  }
+  if (isVector) {
+    const VectorType *VT = dyn_cast<VectorType>(aType);
+    numEle = VT->getNumElements();
+    // We currently cannot support more than 4 elements in a intrinsic and we
+    // cannot support Vec3 types.
+    if (numEle > 4 || numEle == 3) {
+      return false;
+    }
+  }
+  // TODO: Handle vectors.
+  if (isVector) {
+    if (mDebug) {
+      dbgs() << "!!! Vectors are not supported yet!\n";
+    }
+    return false;
+  }
+  Instruction *LHSSrc = NULL, *RHSSrc = NULL;
+  Constant *LHSMask = NULL, *RHSMask = NULL;
+  Constant *LHSShift = NULL, *RHSShift = NULL;
+  Instruction *LHS = dyn_cast<Instruction>(inst->getOperand(0));
+  Instruction *RHS = dyn_cast<Instruction>(inst->getOperand(1));
+  if (!setupBitInsert(LHS, LHSSrc, LHSMask, LHSShift)) {
+    if (mDebug) {
+      dbgs() << "Found an OR Operation that failed setup!\n";
+      inst->dump();
+      if (LHS) { LHS->dump(); }
+      if (LHSSrc) { LHSSrc->dump(); }
+      if (LHSMask) { LHSMask->dump(); }
+      if (LHSShift) { LHSShift->dump(); }
+    }
+    // There was an issue with the setup for BitInsert.
+    return false;
+  }
+  if (!setupBitInsert(RHS, RHSSrc, RHSMask, RHSShift)) {
+    if (mDebug) {
+      dbgs() << "Found an OR Operation that failed setup!\n";
+      inst->dump();
+      if (RHS) { RHS->dump(); }
+      if (RHSSrc) { RHSSrc->dump(); }
+      if (RHSMask) { RHSMask->dump(); }
+      if (RHSShift) { RHSShift->dump(); }
+    }
+    // There was an issue with the setup for BitInsert.
+    return false;
+  }
+  if (mDebug) {
+    dbgs() << "Found an OR operation that can possible be optimized to ubit insert!\n";
+    dbgs() << "Op:        "; inst->dump();
+    dbgs() << "LHS:       "; if (LHS) { LHS->dump(); } else { dbgs() << "(None)\n"; }
+    dbgs() << "LHS Src:   "; if (LHSSrc) { LHSSrc->dump(); } else { dbgs() << "(None)\n"; }
+    dbgs() << "LHS Mask:  "; if (LHSMask) { LHSMask->dump(); } else { dbgs() << "(None)\n"; }
+    dbgs() << "LHS Shift: "; if (LHSShift) { LHSShift->dump(); } else { dbgs() << "(None)\n"; }
+    dbgs() << "RHS:       "; if (RHS) { RHS->dump(); } else { dbgs() << "(None)\n"; }
+    dbgs() << "RHS Src:   "; if (RHSSrc) { RHSSrc->dump(); } else { dbgs() << "(None)\n"; }
+    dbgs() << "RHS Mask:  "; if (RHSMask) { RHSMask->dump(); } else { dbgs() << "(None)\n"; }
+    dbgs() << "RHS Shift: "; if (RHSShift) { RHSShift->dump(); } else { dbgs() << "(None)\n"; }
+  }
+  Constant *offset = NULL;
+  Constant *width = NULL;
+  uint32_t lhsMaskVal = 0, rhsMaskVal = 0;
+  uint32_t lhsShiftVal = 0, rhsShiftVal = 0;
+  uint32_t lhsMaskWidth = 0, rhsMaskWidth = 0;
+  uint32_t lhsMaskOffset = 0, rhsMaskOffset = 0;
+  lhsMaskVal = (LHSMask 
+      ? dyn_cast<ConstantInt>(LHSMask)->getZExtValue() : 0);
+  rhsMaskVal = (RHSMask 
+      ? dyn_cast<ConstantInt>(RHSMask)->getZExtValue() : 0);
+  lhsShiftVal = (LHSShift 
+      ? dyn_cast<ConstantInt>(LHSShift)->getZExtValue() : 0);
+  rhsShiftVal = (RHSShift 
+      ? dyn_cast<ConstantInt>(RHSShift)->getZExtValue() : 0);
+  lhsMaskWidth = lhsMaskVal ? CountPopulation_32(lhsMaskVal) : 32 - lhsShiftVal;
+  rhsMaskWidth = rhsMaskVal ? CountPopulation_32(rhsMaskVal) : 32 - rhsShiftVal;
+  lhsMaskOffset = lhsMaskVal ? CountTrailingZeros_32(lhsMaskVal) : lhsShiftVal;
+  rhsMaskOffset = rhsMaskVal ? CountTrailingZeros_32(rhsMaskVal) : rhsShiftVal;
+  // TODO: Handle the case of A & B | D & ~B(i.e. inverted masks).
+  if ((lhsMaskVal || rhsMaskVal) && !(lhsMaskVal ^ rhsMaskVal)) {
+    return false;
+  }
+  if (lhsMaskOffset >= (rhsMaskWidth + rhsMaskOffset)) {
+    offset = ConstantInt::get(aType, lhsMaskOffset, false);
+    width = ConstantInt::get(aType, lhsMaskWidth, false);
+    RHSSrc = RHS;
+    if (!isMask_32(lhsMaskVal) && !isShiftedMask_32(lhsMaskVal)) {
+      return false;
+    }
+    if (!LHSShift) {
+      LHSSrc = BinaryOperator::Create(Instruction::LShr, LHSSrc, offset,
+          "MaskShr", LHS);
+    } else if (lhsShiftVal != lhsMaskOffset) {
+      LHSSrc = BinaryOperator::Create(Instruction::LShr, LHSSrc, offset,
+          "MaskShr", LHS);
+    }
+    if (mDebug) {
+      dbgs() << "Optimizing LHS!\n";
+    }
+  } else if (rhsMaskOffset >= (lhsMaskWidth + lhsMaskOffset)) {
+    offset = ConstantInt::get(aType, rhsMaskOffset, false);
+    width = ConstantInt::get(aType, rhsMaskWidth, false);
+    LHSSrc = RHSSrc;
+    RHSSrc = LHS;
+    if (!isMask_32(rhsMaskVal) && !isShiftedMask_32(rhsMaskVal)) {
+      return false;
+    }
+    if (!RHSShift) {
+      LHSSrc = BinaryOperator::Create(Instruction::LShr, LHSSrc, offset,
+          "MaskShr", RHS);
+    } else if (rhsShiftVal != rhsMaskOffset) {
+      LHSSrc = BinaryOperator::Create(Instruction::LShr, LHSSrc, offset,
+          "MaskShr", RHS);
+    }
+    if (mDebug) {
+      dbgs() << "Optimizing RHS!\n";
+    }
+  } else {
+    if (mDebug) {
+      dbgs() << "Failed constraint 3!\n";
+    }
+    return false;
+  }
+  if (mDebug) {
+    dbgs() << "Width:  "; if (width) { width->dump(); } else { dbgs() << "(0)\n"; }
+    dbgs() << "Offset: "; if (offset) { offset->dump(); } else { dbgs() << "(0)\n"; }
+    dbgs() << "LHSSrc: "; if (LHSSrc) { LHSSrc->dump(); } else { dbgs() << "(0)\n"; }
+    dbgs() << "RHSSrc: "; if (RHSSrc) { RHSSrc->dump(); } else { dbgs() << "(0)\n"; }
+  }
+  if (!offset || !width) {
+    if (mDebug) {
+      dbgs() << "Either width or offset are NULL, failed detection!\n";
+    }
+    return false;
+  }
+  // Lets create the function signature.
+  std::vector<Type *> callTypes;
+  callTypes.push_back(aType);
+  callTypes.push_back(aType);
+  callTypes.push_back(aType);
+  callTypes.push_back(aType);
+  FunctionType *funcType = FunctionType::get(aType, callTypes, false);
+  std::string name = "__amdil_ubit_insert";
+  if (isVector) { name += "_v" + itostr(numEle) + "u32"; } else { name += "_u32"; }
+  Function *Func = 
+    dyn_cast<Function>(inst->getParent()->getParent()->getParent()->
+        getOrInsertFunction(llvm::StringRef(name), funcType));
+  Value *Operands[4] = {
+    width,
+    offset,
+    LHSSrc,
+    RHSSrc
+  };
+  CallInst *CI = CallInst::Create(Func, Operands, "BitInsertOpt");
+  if (mDebug) {
+    dbgs() << "Old Inst: ";
+    inst->dump();
+    dbgs() << "New Inst: ";
+    CI->dump();
+    dbgs() << "\n\n";
+  }
+  CI->insertBefore(inst);
+  inst->replaceAllUsesWith(CI);
+  return true;
+}
+
+bool 
+AMDGPUPeepholeOpt::optimizeBitExtract(Instruction *inst)  {
+  if (!inst) {
+    return false;
+  }
+  if (!inst->isBinaryOp()) {
+    return false;
+  }
+  if (inst->getOpcode() != Instruction::And) {
+    return false;
+  }
+  if (optLevel == CodeGenOpt::None) {
+    return false;
+  }
+  // We want to do some simple optimizations on Shift right/And patterns. The
+  // basic optimization is to turn (A >> B) & C where A is a 32bit type, B is a
+  // value smaller than 32 and C is a mask. If C is a constant value, then the
+  // following transformation can occur. For signed integers, it turns into the
+  // function call dst = __amdil_ibit_extract(log2(C), B, A) For unsigned
+  // integers, it turns into the function call dst =
+  // __amdil_ubit_extract(log2(C), B, A) The function __amdil_[u|i]bit_extract
+  // can be found in Section 7.9 of the ATI IL spec of the stream SDK for
+  // Evergreen hardware.
+  if (mSTM->device()->getGeneration() == AMDGPUDeviceInfo::HD4XXX) {
+    // This does not work on HD4XXX hardware.
+    return false;
+  }
+  Type *aType = inst->getType();
+  bool isVector = aType->isVectorTy();
+
+  // XXX Support vector types
+  if (isVector) {
+    return false;
+  }
+  int numEle = 1;
+  // This only works on 32bit integers
+  if (aType->getScalarType()
+      != Type::getInt32Ty(inst->getContext())) {
+    return false;
+  }
+  if (isVector) {
+    const VectorType *VT = dyn_cast<VectorType>(aType);
+    numEle = VT->getNumElements();
+    // We currently cannot support more than 4 elements in a intrinsic and we
+    // cannot support Vec3 types.
+    if (numEle > 4 || numEle == 3) {
+      return false;
+    }
+  }
+  BinaryOperator *ShiftInst = dyn_cast<BinaryOperator>(inst->getOperand(0));
+  // If the first operand is not a shift instruction, then we can return as it
+  // doesn't match this pattern.
+  if (!ShiftInst || !ShiftInst->isShift()) {
+    return false;
+  }
+  // If we are a shift left, then we need don't match this pattern.
+  if (ShiftInst->getOpcode() == Instruction::Shl) {
+    return false;
+  }
+  bool isSigned = ShiftInst->isArithmeticShift();
+  Constant *AndMask = dyn_cast<Constant>(inst->getOperand(1));
+  Constant *ShrVal = dyn_cast<Constant>(ShiftInst->getOperand(1));
+  // Lets make sure that the shift value and the and mask are constant integers.
+  if (!AndMask || !ShrVal) {
+    return false;
+  }
+  Constant *newMaskConst;
+  Constant *shiftValConst;
+  if (isVector) {
+    // Handle the vector case
+    std::vector<Constant *> maskVals;
+    std::vector<Constant *> shiftVals;
+    ConstantVector *AndMaskVec = dyn_cast<ConstantVector>(AndMask);
+    ConstantVector *ShrValVec = dyn_cast<ConstantVector>(ShrVal);
+    Type *scalarType = AndMaskVec->getType()->getScalarType();
+    assert(AndMaskVec->getNumOperands() ==
+           ShrValVec->getNumOperands() && "cannot have a "
+           "combination where the number of elements to a "
+           "shift and an and are different!");
+    for (size_t x = 0, y = AndMaskVec->getNumOperands(); x < y; ++x) {
+      ConstantInt *AndCI = dyn_cast<ConstantInt>(AndMaskVec->getOperand(x));
+      ConstantInt *ShiftIC = dyn_cast<ConstantInt>(ShrValVec->getOperand(x));
+      if (!AndCI || !ShiftIC) {
+        return false;
+      }
+      uint32_t maskVal = (uint32_t)AndCI->getZExtValue();
+      if (!isMask_32(maskVal)) {
+        return false;
+      }
+      maskVal = (uint32_t)CountTrailingOnes_32(maskVal);
+      uint32_t shiftVal = (uint32_t)ShiftIC->getZExtValue();
+      // If the mask or shiftval is greater than the bitcount, then break out.
+      if (maskVal >= 32 || shiftVal >= 32) {
+        return false;
+      }
+      // If the mask val is greater than the the number of original bits left
+      // then this optimization is invalid.
+      if (maskVal > (32 - shiftVal)) {
+        return false;
+      }
+      maskVals.push_back(ConstantInt::get(scalarType, maskVal, isSigned));
+      shiftVals.push_back(ConstantInt::get(scalarType, shiftVal, isSigned));
+    }
+    newMaskConst = ConstantVector::get(maskVals);
+    shiftValConst = ConstantVector::get(shiftVals);
+  } else {
+    // Handle the scalar case
+    uint32_t maskVal = (uint32_t)dyn_cast<ConstantInt>(AndMask)->getZExtValue();
+    // This must be a mask value where all lower bits are set to 1 and then any
+    // bit higher is set to 0.
+    if (!isMask_32(maskVal)) {
+      return false;
+    }
+    maskVal = (uint32_t)CountTrailingOnes_32(maskVal);
+    // Count the number of bits set in the mask, this is the width of the
+    // resulting bit set that is extracted from the source value.
+    uint32_t shiftVal = (uint32_t)dyn_cast<ConstantInt>(ShrVal)->getZExtValue();
+    // If the mask or shift val is greater than the bitcount, then break out.
+    if (maskVal >= 32 || shiftVal >= 32) {
+      return false;
+    }
+    // If the mask val is greater than the the number of original bits left then
+    // this optimization is invalid.
+    if (maskVal > (32 - shiftVal)) {
+      return false;
+    }
+    newMaskConst = ConstantInt::get(aType, maskVal, isSigned);
+    shiftValConst = ConstantInt::get(aType, shiftVal, isSigned);
+  }
+  // Lets create the function signature.
+  std::vector<Type *> callTypes;
+  callTypes.push_back(aType);
+  callTypes.push_back(aType);
+  callTypes.push_back(aType);
+  FunctionType *funcType = FunctionType::get(aType, callTypes, false);
+  std::string name = "llvm.AMDGPU.bit.extract.u32";
+  if (isVector) {
+    name += ".v" + itostr(numEle) + "i32";
+  } else {
+    name += ".";
+  }
+  // Lets create the function.
+  Function *Func = 
+    dyn_cast<Function>(inst->getParent()->getParent()->getParent()->
+                       getOrInsertFunction(llvm::StringRef(name), funcType));
+  Value *Operands[3] = {
+    ShiftInst->getOperand(0),
+    shiftValConst,
+    newMaskConst
+  };
+  // Lets create the Call with the operands
+  CallInst *CI = CallInst::Create(Func, Operands, "ByteExtractOpt");
+  CI->setDoesNotAccessMemory();
+  CI->insertBefore(inst);
+  inst->replaceAllUsesWith(CI);
+  return true;
+}
+
+bool
+AMDGPUPeepholeOpt::expandBFI(CallInst *CI) {
+  if (!CI) {
+    return false;
+  }
+  Value *LHS = CI->getOperand(CI->getNumOperands() - 1);
+  if (!LHS->getName().startswith("__amdil_bfi")) {
+    return false;
+  }
+  Type* type = CI->getOperand(0)->getType();
+  Constant *negOneConst = NULL;
+  if (type->isVectorTy()) {
+    std::vector<Constant *> negOneVals;
+    negOneConst = ConstantInt::get(CI->getContext(), 
+        APInt(32, StringRef("-1"), 10));
+    for (size_t x = 0,
+        y = dyn_cast<VectorType>(type)->getNumElements(); x < y; ++x) {
+      negOneVals.push_back(negOneConst);
+    }
+    negOneConst = ConstantVector::get(negOneVals);
+  } else {
+    negOneConst = ConstantInt::get(CI->getContext(), 
+        APInt(32, StringRef("-1"), 10));
+  }
+  // __amdil_bfi => (A & B) | (~A & C)
+  BinaryOperator *lhs = 
+    BinaryOperator::Create(Instruction::And, CI->getOperand(0),
+        CI->getOperand(1), "bfi_and", CI);
+  BinaryOperator *rhs =
+    BinaryOperator::Create(Instruction::Xor, CI->getOperand(0), negOneConst,
+        "bfi_not", CI);
+  rhs = BinaryOperator::Create(Instruction::And, rhs, CI->getOperand(2),
+      "bfi_and", CI);
+  lhs = BinaryOperator::Create(Instruction::Or, lhs, rhs, "bfi_or", CI);
+  CI->replaceAllUsesWith(lhs);
+  return true;
+}
+
+bool
+AMDGPUPeepholeOpt::expandBFM(CallInst *CI) {
+  if (!CI) {
+    return false;
+  }
+  Value *LHS = CI->getOperand(CI->getNumOperands() - 1);
+  if (!LHS->getName().startswith("__amdil_bfm")) {
+    return false;
+  }
+  // __amdil_bfm => ((1 << (src0 & 0x1F)) - 1) << (src1 & 0x1f)
+  Constant *newMaskConst = NULL;
+  Constant *newShiftConst = NULL;
+  Type* type = CI->getOperand(0)->getType();
+  if (type->isVectorTy()) {
+    std::vector<Constant*> newMaskVals, newShiftVals;
+    newMaskConst = ConstantInt::get(Type::getInt32Ty(*mCTX), 0x1F);
+    newShiftConst = ConstantInt::get(Type::getInt32Ty(*mCTX), 1);
+    for (size_t x = 0,
+        y = dyn_cast<VectorType>(type)->getNumElements(); x < y; ++x) {
+      newMaskVals.push_back(newMaskConst);
+      newShiftVals.push_back(newShiftConst);
+    }
+    newMaskConst = ConstantVector::get(newMaskVals);
+    newShiftConst = ConstantVector::get(newShiftVals);
+  } else {
+    newMaskConst = ConstantInt::get(Type::getInt32Ty(*mCTX), 0x1F);
+    newShiftConst = ConstantInt::get(Type::getInt32Ty(*mCTX), 1);
+  }
+  BinaryOperator *lhs =
+    BinaryOperator::Create(Instruction::And, CI->getOperand(0),
+        newMaskConst, "bfm_mask", CI);
+  lhs = BinaryOperator::Create(Instruction::Shl, newShiftConst,
+      lhs, "bfm_shl", CI);
+  lhs = BinaryOperator::Create(Instruction::Sub, lhs,
+      newShiftConst, "bfm_sub", CI);
+  BinaryOperator *rhs =
+    BinaryOperator::Create(Instruction::And, CI->getOperand(1),
+        newMaskConst, "bfm_mask", CI);
+  lhs = BinaryOperator::Create(Instruction::Shl, lhs, rhs, "bfm_shl", CI);
+  CI->replaceAllUsesWith(lhs);
+  return true;
+}
+
+bool
+AMDGPUPeepholeOpt::instLevelOptimizations(BasicBlock::iterator *bbb)  {
+  Instruction *inst = (*bbb);
+  if (optimizeCallInst(bbb)) {
+    return true;
+  }
+  if (optimizeBitExtract(inst)) {
+    return false;
+  }
+  if (optimizeBitInsert(inst)) {
+    return false;
+  }
+  if (correctMisalignedMemOp(inst)) {
+    return false;
+  }
+  return false;
+}
+bool
+AMDGPUPeepholeOpt::correctMisalignedMemOp(Instruction *inst) {
+  LoadInst *linst = dyn_cast<LoadInst>(inst);
+  StoreInst *sinst = dyn_cast<StoreInst>(inst);
+  unsigned alignment;
+  Type* Ty = inst->getType();
+  if (linst) {
+    alignment = linst->getAlignment();
+    Ty = inst->getType();
+  } else if (sinst) {
+    alignment = sinst->getAlignment();
+    Ty = sinst->getValueOperand()->getType();
+  } else {
+    return false;
+  }
+  unsigned size = getTypeSize(Ty);
+  if (size == alignment || size < alignment) {
+    return false;
+  }
+  if (!Ty->isStructTy()) {
+    return false;
+  }
+  if (alignment < 4) {
+    if (linst) {
+      linst->setAlignment(0);
+      return true;
+    } else if (sinst) {
+      sinst->setAlignment(0);
+      return true;
+    }
+  }
+  return false;
+}
+bool 
+AMDGPUPeepholeOpt::isSigned24BitOps(CallInst *CI)  {
+  if (!CI) {
+    return false;
+  }
+  Value *LHS = CI->getOperand(CI->getNumOperands() - 1);
+  std::string namePrefix = LHS->getName().substr(0, 14);
+  if (namePrefix != "__amdil_imad24" && namePrefix != "__amdil_imul24"
+      && namePrefix != "__amdil__imul24_high") {
+    return false;
+  }
+  if (mSTM->device()->usesHardware(AMDGPUDeviceInfo::Signed24BitOps)) {
+    return false;
+  }
+  return true;
+}
+
+void 
+AMDGPUPeepholeOpt::expandSigned24BitOps(CallInst *CI)  {
+  assert(isSigned24BitOps(CI) && "Must be a "
+      "signed 24 bit operation to call this function!");
+  Value *LHS = CI->getOperand(CI->getNumOperands()-1);
+  // On 7XX and 8XX we do not have signed 24bit, so we need to
+  // expand it to the following:
+  // imul24 turns into 32bit imul
+  // imad24 turns into 32bit imad
+  // imul24_high turns into 32bit imulhigh
+  if (LHS->getName().substr(0, 14) == "__amdil_imad24") {
+    Type *aType = CI->getOperand(0)->getType();
+    bool isVector = aType->isVectorTy();
+    int numEle = isVector ? dyn_cast<VectorType>(aType)->getNumElements() : 1;
+    std::vector<Type*> callTypes;
+    callTypes.push_back(CI->getOperand(0)->getType());
+    callTypes.push_back(CI->getOperand(1)->getType());
+    callTypes.push_back(CI->getOperand(2)->getType());
+    FunctionType *funcType =
+      FunctionType::get(CI->getOperand(0)->getType(), callTypes, false);
+    std::string name = "__amdil_imad";
+    if (isVector) {
+      name += "_v" + itostr(numEle) + "i32";
+    } else {
+      name += "_i32";
+    }
+    Function *Func = dyn_cast<Function>(
+                       CI->getParent()->getParent()->getParent()->
+                       getOrInsertFunction(llvm::StringRef(name), funcType));
+    Value *Operands[3] = {
+      CI->getOperand(0),
+      CI->getOperand(1),
+      CI->getOperand(2)
+    };
+    CallInst *nCI = CallInst::Create(Func, Operands, "imad24");
+    nCI->insertBefore(CI);
+    CI->replaceAllUsesWith(nCI);
+  } else if (LHS->getName().substr(0, 14) == "__amdil_imul24") {
+    BinaryOperator *mulOp =
+      BinaryOperator::Create(Instruction::Mul, CI->getOperand(0),
+          CI->getOperand(1), "imul24", CI);
+    CI->replaceAllUsesWith(mulOp);
+  } else if (LHS->getName().substr(0, 19) == "__amdil_imul24_high") {
+    Type *aType = CI->getOperand(0)->getType();
+
+    bool isVector = aType->isVectorTy();
+    int numEle = isVector ? dyn_cast<VectorType>(aType)->getNumElements() : 1;
+    std::vector<Type*> callTypes;
+    callTypes.push_back(CI->getOperand(0)->getType());
+    callTypes.push_back(CI->getOperand(1)->getType());
+    FunctionType *funcType =
+      FunctionType::get(CI->getOperand(0)->getType(), callTypes, false);
+    std::string name = "__amdil_imul_high";
+    if (isVector) {
+      name += "_v" + itostr(numEle) + "i32";
+    } else {
+      name += "_i32";
+    }
+    Function *Func = dyn_cast<Function>(
+                       CI->getParent()->getParent()->getParent()->
+                       getOrInsertFunction(llvm::StringRef(name), funcType));
+    Value *Operands[2] = {
+      CI->getOperand(0),
+      CI->getOperand(1)
+    };
+    CallInst *nCI = CallInst::Create(Func, Operands, "imul24_high");
+    nCI->insertBefore(CI);
+    CI->replaceAllUsesWith(nCI);
+  }
+}
+
+bool 
+AMDGPUPeepholeOpt::isRWGLocalOpt(CallInst *CI)  {
+  return (CI != NULL
+          && CI->getOperand(CI->getNumOperands() - 1)->getName() 
+          == "__amdil_get_local_size_int");
+}
+
+bool 
+AMDGPUPeepholeOpt::convertAccurateDivide(CallInst *CI)  {
+  if (!CI) {
+    return false;
+  }
+  if (mSTM->device()->getGeneration() == AMDGPUDeviceInfo::HD6XXX
+      && (mSTM->getDeviceName() == "cayman")) {
+    return false;
+  }
+  return CI->getOperand(CI->getNumOperands() - 1)->getName().substr(0, 20) 
+      == "__amdil_improved_div";
+}
+
+void 
+AMDGPUPeepholeOpt::expandAccurateDivide(CallInst *CI)  {
+  assert(convertAccurateDivide(CI)
+         && "expanding accurate divide can only happen if it is expandable!");
+  BinaryOperator *divOp =
+    BinaryOperator::Create(Instruction::FDiv, CI->getOperand(0),
+                           CI->getOperand(1), "fdiv32", CI);
+  CI->replaceAllUsesWith(divOp);
+}
+
+bool
+AMDGPUPeepholeOpt::propagateSamplerInst(CallInst *CI) {
+  if (optLevel != CodeGenOpt::None) {
+    return false;
+  }
+
+  if (!CI) {
+    return false;
+  }
+
+  unsigned funcNameIdx = 0;
+  funcNameIdx = CI->getNumOperands() - 1;
+  StringRef calleeName = CI->getOperand(funcNameIdx)->getName();
+  if (calleeName != "__amdil_image2d_read_norm"
+   && calleeName != "__amdil_image2d_read_unnorm"
+   && calleeName != "__amdil_image3d_read_norm"
+   && calleeName != "__amdil_image3d_read_unnorm") {
+    return false;
+  }
+
+  unsigned samplerIdx = 2;
+  samplerIdx = 1;
+  Value *sampler = CI->getOperand(samplerIdx);
+  LoadInst *lInst = dyn_cast<LoadInst>(sampler);
+  if (!lInst) {
+    return false;
+  }
+
+  if (lInst->getPointerAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS) {
+    return false;
+  }
+
+  GlobalVariable *gv = dyn_cast<GlobalVariable>(lInst->getPointerOperand());
+  // If we are loading from what is not a global value, then we
+  // fail and return.
+  if (!gv) {
+    return false;
+  }
+
+  // If we don't have an initializer or we have an initializer and
+  // the initializer is not a 32bit integer, we fail.
+  if (!gv->hasInitializer() 
+      || !gv->getInitializer()->getType()->isIntegerTy(32)) {
+      return false;
+  }
+
+  // Now that we have the global variable initializer, lets replace
+  // all uses of the load instruction with the samplerVal and
+  // reparse the __amdil_is_constant() function.
+  Constant *samplerVal = gv->getInitializer();
+  lInst->replaceAllUsesWith(samplerVal);
+  return true;
+}
+
+bool 
+AMDGPUPeepholeOpt::doInitialization(Module &M)  {
+  return false;
+}
+
+bool 
+AMDGPUPeepholeOpt::doFinalization(Module &M)  {
+  return false;
+}
+
+void 
+AMDGPUPeepholeOpt::getAnalysisUsage(AnalysisUsage &AU) const  {
+  AU.addRequired<MachineFunctionAnalysis>();
+  FunctionPass::getAnalysisUsage(AU);
+  AU.setPreservesAll();
+}
+
+size_t AMDGPUPeepholeOpt::getTypeSize(Type * const T, bool dereferencePtr) {
+  size_t size = 0;
+  if (!T) {
+    return size;
+  }
+  switch (T->getTypeID()) {
+  case Type::X86_FP80TyID:
+  case Type::FP128TyID:
+  case Type::PPC_FP128TyID:
+  case Type::LabelTyID:
+    assert(0 && "These types are not supported by this backend");
+  default:
+  case Type::FloatTyID:
+  case Type::DoubleTyID:
+    size = T->getPrimitiveSizeInBits() >> 3;
+    break;
+  case Type::PointerTyID:
+    size = getTypeSize(dyn_cast<PointerType>(T), dereferencePtr);
+    break;
+  case Type::IntegerTyID:
+    size = getTypeSize(dyn_cast<IntegerType>(T), dereferencePtr);
+    break;
+  case Type::StructTyID:
+    size = getTypeSize(dyn_cast<StructType>(T), dereferencePtr);
+    break;
+  case Type::ArrayTyID:
+    size = getTypeSize(dyn_cast<ArrayType>(T), dereferencePtr);
+    break;
+  case Type::FunctionTyID:
+    size = getTypeSize(dyn_cast<FunctionType>(T), dereferencePtr);
+    break;
+  case Type::VectorTyID:
+    size = getTypeSize(dyn_cast<VectorType>(T), dereferencePtr);
+    break;
+  };
+  return size;
+}
+
+size_t AMDGPUPeepholeOpt::getTypeSize(StructType * const ST,
+    bool dereferencePtr) {
+  size_t size = 0;
+  if (!ST) {
+    return size;
+  }
+  Type *curType;
+  StructType::element_iterator eib;
+  StructType::element_iterator eie;
+  for (eib = ST->element_begin(), eie = ST->element_end(); eib != eie; ++eib) {
+    curType = *eib;
+    size += getTypeSize(curType, dereferencePtr);
+  }
+  return size;
+}
+
+size_t AMDGPUPeepholeOpt::getTypeSize(IntegerType * const IT,
+    bool dereferencePtr) {
+  return IT ? (IT->getBitWidth() >> 3) : 0;
+}
+
+size_t AMDGPUPeepholeOpt::getTypeSize(FunctionType * const FT,
+    bool dereferencePtr) {
+    assert(0 && "Should not be able to calculate the size of an function type");
+    return 0;
+}
+
+size_t AMDGPUPeepholeOpt::getTypeSize(ArrayType * const AT,
+    bool dereferencePtr) {
+  return (size_t)(AT ? (getTypeSize(AT->getElementType(),
+                                    dereferencePtr) * AT->getNumElements())
+                     : 0);
+}
+
+size_t AMDGPUPeepholeOpt::getTypeSize(VectorType * const VT,
+    bool dereferencePtr) {
+  return VT ? (VT->getBitWidth() >> 3) : 0;
+}
+
+size_t AMDGPUPeepholeOpt::getTypeSize(PointerType * const PT,
+    bool dereferencePtr) {
+  if (!PT) {
+    return 0;
+  }
+  Type *CT = PT->getElementType();
+  if (CT->getTypeID() == Type::StructTyID &&
+      PT->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) {
+    return getTypeSize(dyn_cast<StructType>(CT));
+  } else if (dereferencePtr) {
+    size_t size = 0;
+    for (size_t x = 0, y = PT->getNumContainedTypes(); x < y; ++x) {
+      size += getTypeSize(PT->getContainedType(x), dereferencePtr);
+    }
+    return size;
+  } else {
+    return 4;
+  }
+}
+
+size_t AMDGPUPeepholeOpt::getTypeSize(OpaqueType * const OT,
+    bool dereferencePtr) {
+  //assert(0 && "Should not be able to calculate the size of an opaque type");
+  return 4;
+}
diff --git a/lib/Target/R600/AMDILRegisterInfo.td b/lib/Target/R600/AMDILRegisterInfo.td
new file mode 100644
index 0000000000..b9d033432e
--- /dev/null
+++ b/lib/Target/R600/AMDILRegisterInfo.td
@@ -0,0 +1,107 @@
+//===- AMDILRegisterInfo.td - AMDIL Register defs ----------*- tablegen -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+//  Declarations that describe the AMDIL register file
+//
+//===----------------------------------------------------------------------===//
+
+class AMDILReg<bits<16> num, string n> : Register<n> {
+  field bits<16> Value;
+  let Value = num;
+  let Namespace = "AMDGPU";
+}
+
+// We will start with 8 registers for each class before expanding to more
+// Since the swizzle is added based on the register class, we can leave it
+// off here and just specify different registers for different register classes
+def R1 : AMDILReg<1, "r1">, DwarfRegNum<[1]>;
+def R2 : AMDILReg<2, "r2">, DwarfRegNum<[2]>;
+def R3 : AMDILReg<3, "r3">, DwarfRegNum<[3]>;
+def R4 : AMDILReg<4, "r4">, DwarfRegNum<[4]>;
+def R5 : AMDILReg<5, "r5">, DwarfRegNum<[5]>;
+def R6 : AMDILReg<6, "r6">, DwarfRegNum<[6]>;
+def R7 : AMDILReg<7, "r7">, DwarfRegNum<[7]>;
+def R8 : AMDILReg<8, "r8">, DwarfRegNum<[8]>;
+def R9 : AMDILReg<9, "r9">, DwarfRegNum<[9]>;
+def R10 : AMDILReg<10, "r10">, DwarfRegNum<[10]>;
+def R11 : AMDILReg<11, "r11">, DwarfRegNum<[11]>;
+def R12 : AMDILReg<12, "r12">, DwarfRegNum<[12]>;
+def R13 : AMDILReg<13, "r13">, DwarfRegNum<[13]>;
+def R14 : AMDILReg<14, "r14">, DwarfRegNum<[14]>;
+def R15 : AMDILReg<15, "r15">, DwarfRegNum<[15]>;
+def R16 : AMDILReg<16, "r16">, DwarfRegNum<[16]>;
+def R17 : AMDILReg<17, "r17">, DwarfRegNum<[17]>;
+def R18 : AMDILReg<18, "r18">, DwarfRegNum<[18]>;
+def R19 : AMDILReg<19, "r19">, DwarfRegNum<[19]>;
+def R20 : AMDILReg<20, "r20">, DwarfRegNum<[20]>;
+
+// All registers between 1000 and 1024 are reserved and cannot be used
+// unless commented in this section
+// r1021-r1025 are used to dynamically calculate the local/group/thread/region/region_local ID's
+// r1020 is used to hold the frame index for local arrays
+// r1019 is used to hold the dynamic stack allocation pointer
+// r1018 is used as a temporary register for handwritten code
+// r1017 is used as a temporary register for handwritten code
+// r1016 is used as a temporary register for load/store code
+// r1015 is used as a temporary register for data segment offset
+// r1014 is used as a temporary register for store code
+// r1013 is used as the section data pointer register
+// r1012-r1010 and r1001-r1008 are used for temporary I/O registers
+// r1009 is used as the frame pointer register
+// r999 is used as the mem register.
+// r998 is used as the return address register.
+//def R1025 : AMDILReg<1025, "r1025">, DwarfRegNum<[1025]>;
+//def R1024 : AMDILReg<1024, "r1024">, DwarfRegNum<[1024]>;
+//def R1023 : AMDILReg<1023, "r1023">, DwarfRegNum<[1023]>;
+//def R1022 : AMDILReg<1022, "r1022">, DwarfRegNum<[1022]>;
+//def R1021 : AMDILReg<1021, "r1021">, DwarfRegNum<[1021]>;
+//def R1020 : AMDILReg<1020, "r1020">, DwarfRegNum<[1020]>;
+def SP : AMDILReg<1019, "r1019">, DwarfRegNum<[1019]>;
+def T1 : AMDILReg<1018, "r1018">, DwarfRegNum<[1018]>;
+def T2 : AMDILReg<1017, "r1017">, DwarfRegNum<[1017]>;
+def T3 : AMDILReg<1016, "r1016">, DwarfRegNum<[1016]>;
+def T4 : AMDILReg<1015, "r1015">, DwarfRegNum<[1015]>;
+def T5 : AMDILReg<1014, "r1014">, DwarfRegNum<[1014]>;
+def SDP : AMDILReg<1013, "r1013">, DwarfRegNum<[1013]>;
+def R1012: AMDILReg<1012, "r1012">, DwarfRegNum<[1012]>;
+def R1011: AMDILReg<1011, "r1011">, DwarfRegNum<[1011]>;
+def R1010: AMDILReg<1010, "r1010">, DwarfRegNum<[1010]>;
+def DFP : AMDILReg<1009, "r1009">, DwarfRegNum<[1009]>;
+def R1008: AMDILReg<1008, "r1008">, DwarfRegNum<[1008]>;
+def R1007: AMDILReg<1007, "r1007">, DwarfRegNum<[1007]>;
+def R1006: AMDILReg<1006, "r1006">, DwarfRegNum<[1006]>;
+def R1005: AMDILReg<1005, "r1005">, DwarfRegNum<[1005]>;
+def R1004: AMDILReg<1004, "r1004">, DwarfRegNum<[1004]>;
+def R1003: AMDILReg<1003, "r1003">, DwarfRegNum<[1003]>;
+def R1002: AMDILReg<1002, "r1002">, DwarfRegNum<[1002]>;
+def R1001: AMDILReg<1001, "r1001">, DwarfRegNum<[1001]>;
+def MEM : AMDILReg<999, "mem">, DwarfRegNum<[999]>;
+def RA : AMDILReg<998, "r998">, DwarfRegNum<[998]>;
+def FP : AMDILReg<997, "r997">, DwarfRegNum<[997]>;
+def GPRI16 : RegisterClass<"AMDGPU", [i16], 16,
+  (add (sequence "R%u", 1, 20), RA, SP, T1, T2, T3, T4, T5, SDP, R1010, R1011, R1001, R1002, R1003, R1004, R1005, R1006, R1007, R1008, MEM, R1012)> {
+        let AltOrders = [(add (sequence "R%u", 1, 20))];
+        let AltOrderSelect = [{
+          return 1;
+        }];
+    }
+def GPRI32 : RegisterClass<"AMDGPU", [i32], 32,
+  (add (sequence "R%u", 1, 20), RA, SP, T1, T2, T3, T4, T5, SDP, R1010, R1011, R1001, R1002, R1003, R1004, R1005, R1006, R1007, R1008, MEM, R1012)> {
+        let AltOrders = [(add (sequence "R%u", 1, 20))];
+        let AltOrderSelect = [{
+          return 1;
+        }];
+    }
+def GPRF32 : RegisterClass<"AMDGPU", [f32], 32,
+  (add (sequence "R%u", 1, 20), RA, SP, T1, T2, T3, T4, T5, SDP, R1010, R1011, R1001, R1002, R1003, R1004, R1005, R1006, R1007, R1008, MEM, R1012)> {
+        let AltOrders = [(add (sequence "R%u", 1, 20))];
+        let AltOrderSelect = [{
+          return 1;
+        }];
+    }
diff --git a/lib/Target/R600/AMDILSIDevice.cpp b/lib/Target/R600/AMDILSIDevice.cpp
new file mode 100644
index 0000000000..7c2710f1b2
--- /dev/null
+++ b/lib/Target/R600/AMDILSIDevice.cpp
@@ -0,0 +1,45 @@
+//===-- AMDILSIDevice.cpp - Device Info for Southern Islands GPUs ---------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//==-----------------------------------------------------------------------===//
+#include "AMDILSIDevice.h"
+#include "AMDILEvergreenDevice.h"
+#include "AMDILNIDevice.h"
+#include "AMDGPUSubtarget.h"
+
+using namespace llvm;
+
+AMDGPUSIDevice::AMDGPUSIDevice(AMDGPUSubtarget *ST)
+  : AMDGPUEvergreenDevice(ST) {
+}
+AMDGPUSIDevice::~AMDGPUSIDevice() {
+}
+
+size_t
+AMDGPUSIDevice::getMaxLDSSize() const {
+  if (usesHardware(AMDGPUDeviceInfo::LocalMem)) {
+    return MAX_LDS_SIZE_900;
+  } else {
+    return 0;
+  }
+}
+
+uint32_t
+AMDGPUSIDevice::getGeneration() const {
+  return AMDGPUDeviceInfo::HD7XXX;
+}
+
+std::string
+AMDGPUSIDevice::getDataLayout() const {
+  return std::string("e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16"
+      "-i32:32:32-i64:64:64-f32:32:32-f64:64:64-f80:32:32"
+      "-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64"
+      "-v96:128:128-v128:128:128-v192:256:256-v256:256:256"
+      "-v512:512:512-v1024:1024:1024-v2048:2048:2048"
+      "-n8:16:32:64");
+}
diff --git a/lib/Target/R600/AMDILSIDevice.h b/lib/Target/R600/AMDILSIDevice.h
new file mode 100644
index 0000000000..5b2cb25022
--- /dev/null
+++ b/lib/Target/R600/AMDILSIDevice.h
@@ -0,0 +1,39 @@
+//===------- AMDILSIDevice.h - Define SI Device for AMDIL -*- C++ -*------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Interface for the subtarget data classes.
+///
+/// This file will define the interface that each generation needs to
+/// implement in order to correctly answer queries on the capabilities of the
+/// specific hardware.
+//===---------------------------------------------------------------------===//
+#ifndef AMDILSIDEVICE_H
+#define AMDILSIDEVICE_H
+#include "AMDILEvergreenDevice.h"
+
+namespace llvm {
+class AMDGPUSubtarget;
+//===---------------------------------------------------------------------===//
+// SI generation of devices and their respective sub classes
+//===---------------------------------------------------------------------===//
+
+/// \brief The AMDGPUSIDevice is the base class for all Southern Island series
+/// of cards.
+class AMDGPUSIDevice : public AMDGPUEvergreenDevice {
+public:
+  AMDGPUSIDevice(AMDGPUSubtarget*);
+  virtual ~AMDGPUSIDevice();
+  virtual size_t getMaxLDSSize() const;
+  virtual uint32_t getGeneration() const;
+  virtual std::string getDataLayout() const;
+};
+
+} // namespace llvm
+#endif // AMDILSIDEVICE_H
diff --git a/lib/Target/R600/CMakeLists.txt b/lib/Target/R600/CMakeLists.txt
new file mode 100644
index 0000000000..757b733522
--- /dev/null
+++ b/lib/Target/R600/CMakeLists.txt
@@ -0,0 +1,54 @@
+set(LLVM_TARGET_DEFINITIONS AMDGPU.td)
+
+tablegen(LLVM AMDGPUGenRegisterInfo.inc -gen-register-info)
+tablegen(LLVM AMDGPUGenInstrInfo.inc -gen-instr-info)
+tablegen(LLVM AMDGPUGenDAGISel.inc -gen-dag-isel)
+tablegen(LLVM AMDGPUGenCallingConv.inc -gen-callingconv)
+tablegen(LLVM AMDGPUGenSubtargetInfo.inc -gen-subtarget)
+tablegen(LLVM AMDGPUGenIntrinsics.inc -gen-tgt-intrinsic)
+tablegen(LLVM AMDGPUGenMCCodeEmitter.inc -gen-emitter -mc-emitter)
+tablegen(LLVM AMDGPUGenDFAPacketizer.inc -gen-dfa-packetizer)
+tablegen(LLVM AMDGPUGenAsmWriter.inc -gen-asm-writer)
+add_public_tablegen_target(AMDGPUCommonTableGen)
+
+add_llvm_target(R600CodeGen
+  AMDIL7XXDevice.cpp
+  AMDILCFGStructurizer.cpp
+  AMDILDevice.cpp
+  AMDILDeviceInfo.cpp
+  AMDILEvergreenDevice.cpp
+  AMDILFrameLowering.cpp
+  AMDILIntrinsicInfo.cpp
+  AMDILISelDAGToDAG.cpp
+  AMDILISelLowering.cpp
+  AMDILNIDevice.cpp
+  AMDILPeepholeOptimizer.cpp
+  AMDILSIDevice.cpp
+  AMDGPUAsmPrinter.cpp
+  AMDGPUMCInstLower.cpp
+  AMDGPUSubtarget.cpp
+  AMDGPUTargetMachine.cpp
+  AMDGPUISelLowering.cpp
+  AMDGPUConvertToISA.cpp
+  AMDGPUInstrInfo.cpp
+  AMDGPURegisterInfo.cpp
+  R600ExpandSpecialInstrs.cpp
+  R600InstrInfo.cpp
+  R600ISelLowering.cpp
+  R600MachineFunctionInfo.cpp
+  R600RegisterInfo.cpp
+  SIAssignInterpRegs.cpp
+  SIInstrInfo.cpp
+  SIISelLowering.cpp
+  SILowerLiteralConstants.cpp
+  SILowerControlFlow.cpp
+  SIMachineFunctionInfo.cpp
+  SIRegisterInfo.cpp
+  SIFixSGPRLiveness.cpp
+  )
+
+add_dependencies(LLVMR600CodeGen intrinsics_gen)
+
+add_subdirectory(InstPrinter)
+add_subdirectory(TargetInfo)
+add_subdirectory(MCTargetDesc)
diff --git a/lib/Target/R600/InstPrinter/AMDGPUInstPrinter.cpp b/lib/Target/R600/InstPrinter/AMDGPUInstPrinter.cpp
new file mode 100644
index 0000000000..e6c550b5ac
--- /dev/null
+++ b/lib/Target/R600/InstPrinter/AMDGPUInstPrinter.cpp
@@ -0,0 +1,132 @@
+//===-- AMDGPUInstPrinter.cpp - AMDGPU MC Inst -> ASM ---------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+// \file
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUInstPrinter.h"
+#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
+#include "llvm/MC/MCInst.h"
+
+using namespace llvm;
+
+void AMDGPUInstPrinter::printInst(const MCInst *MI, raw_ostream &OS,
+                             StringRef Annot) {
+  printInstruction(MI, OS);
+
+  printAnnotation(OS, Annot);
+}
+
+void AMDGPUInstPrinter::printOperand(const MCInst *MI, unsigned OpNo,
+                                     raw_ostream &O) {
+
+  const MCOperand &Op = MI->getOperand(OpNo);
+  if (Op.isReg()) {
+    switch (Op.getReg()) {
+    // This is the default predicate state, so we don't need to print it.
+    case AMDGPU::PRED_SEL_OFF: break;
+    default: O << getRegisterName(Op.getReg()); break;
+    }
+  } else if (Op.isImm()) {
+    O << Op.getImm();
+  } else if (Op.isFPImm()) {
+    O << Op.getFPImm();
+  } else {
+    assert(!"unknown operand type in printOperand");
+  }
+}
+
+void AMDGPUInstPrinter::printMemOperand(const MCInst *MI, unsigned OpNo,
+                                        raw_ostream &O) {
+  printOperand(MI, OpNo, O);
+  O  << ", ";
+  printOperand(MI, OpNo + 1, O);
+}
+
+void AMDGPUInstPrinter::printIfSet(const MCInst *MI, unsigned OpNo,
+                                    raw_ostream &O, StringRef Asm) {
+  const MCOperand &Op = MI->getOperand(OpNo);
+  assert(Op.isImm());
+  if (Op.getImm() == 1) {
+    O << Asm;
+  }
+}
+
+void AMDGPUInstPrinter::printAbs(const MCInst *MI, unsigned OpNo,
+                                 raw_ostream &O) {
+  printIfSet(MI, OpNo, O, "|");
+}
+
+void AMDGPUInstPrinter::printClamp(const MCInst *MI, unsigned OpNo,
+                                   raw_ostream &O) {
+  printIfSet(MI, OpNo, O, "_SAT");
+}
+
+void AMDGPUInstPrinter::printLiteral(const MCInst *MI, unsigned OpNo,
+                                     raw_ostream &O) {
+  union Literal {
+    float f;
+    int32_t i;
+  } L;
+
+  L.i = MI->getOperand(OpNo).getImm();
+  O << L.i << "(" << L.f << ")";
+}
+
+void AMDGPUInstPrinter::printLast(const MCInst *MI, unsigned OpNo,
+                                  raw_ostream &O) {
+  printIfSet(MI, OpNo, O, " *");
+}
+
+void AMDGPUInstPrinter::printNeg(const MCInst *MI, unsigned OpNo,
+                                 raw_ostream &O) {
+  printIfSet(MI, OpNo, O, "-");
+}
+
+void AMDGPUInstPrinter::printOMOD(const MCInst *MI, unsigned OpNo,
+                                  raw_ostream &O) {
+  switch (MI->getOperand(OpNo).getImm()) {
+  default: break;
+  case 1:
+    O << " * 2.0";
+    break;
+  case 2:
+    O << " * 4.0";
+    break;
+  case 3:
+    O << " / 2.0";
+    break;
+  }
+}
+
+void AMDGPUInstPrinter::printRel(const MCInst *MI, unsigned OpNo,
+                                 raw_ostream &O) {
+  const MCOperand &Op = MI->getOperand(OpNo);
+  if (Op.getImm() != 0) {
+    O << " + " << Op.getImm();
+  }
+}
+
+void AMDGPUInstPrinter::printUpdateExecMask(const MCInst *MI, unsigned OpNo,
+                                            raw_ostream &O) {
+  printIfSet(MI, OpNo, O, "ExecMask,");
+}
+
+void AMDGPUInstPrinter::printUpdatePred(const MCInst *MI, unsigned OpNo,
+                                        raw_ostream &O) {
+  printIfSet(MI, OpNo, O, "Pred,");
+}
+
+void AMDGPUInstPrinter::printWrite(const MCInst *MI, unsigned OpNo,
+                                       raw_ostream &O) {
+  const MCOperand &Op = MI->getOperand(OpNo);
+  if (Op.getImm() == 0) {
+    O << " (MASKED)";
+  }
+}
+
+#include "AMDGPUGenAsmWriter.inc"
diff --git a/lib/Target/R600/InstPrinter/AMDGPUInstPrinter.h b/lib/Target/R600/InstPrinter/AMDGPUInstPrinter.h
new file mode 100644
index 0000000000..96e0e46f8a
--- /dev/null
+++ b/lib/Target/R600/InstPrinter/AMDGPUInstPrinter.h
@@ -0,0 +1,52 @@
+//===-- AMDGPUInstPrinter.h - AMDGPU MC Inst -> ASM interface ---*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+//===----------------------------------------------------------------------===//
+
+#ifndef AMDGPUINSTPRINTER_H
+#define AMDGPUINSTPRINTER_H
+
+#include "llvm/ADT/StringRef.h"
+#include "llvm/MC/MCInstPrinter.h"
+#include "llvm/Support/raw_ostream.h"
+
+namespace llvm {
+
+class AMDGPUInstPrinter : public MCInstPrinter {
+public:
+  AMDGPUInstPrinter(const MCAsmInfo &MAI, const MCInstrInfo &MII,
+                     const MCRegisterInfo &MRI)
+    : MCInstPrinter(MAI, MII, MRI) {}
+
+  //Autogenerated by tblgen
+  void printInstruction(const MCInst *MI, raw_ostream &O);
+  static const char *getRegisterName(unsigned RegNo);
+
+  virtual void printInst(const MCInst *MI, raw_ostream &O, StringRef Annot);
+
+private:
+  void printOperand(const MCInst *MI, unsigned OpNo, raw_ostream &O);
+  void printMemOperand(const MCInst *MI, unsigned OpNo, raw_ostream &O);
+  void printIfSet(const MCInst *MI, unsigned OpNo, raw_ostream &O, StringRef Asm);
+  void printAbs(const MCInst *MI, unsigned OpNo, raw_ostream &O);
+  void printClamp(const MCInst *MI, unsigned OpNo, raw_ostream &O);
+  void printLiteral(const MCInst *MI, unsigned OpNo, raw_ostream &O);
+  void printLast(const MCInst *MI, unsigned OpNo, raw_ostream &O);
+  void printNeg(const MCInst *MI, unsigned OpNo, raw_ostream &O);
+  void printOMOD(const MCInst *MI, unsigned OpNo, raw_ostream &O);
+  void printRel(const MCInst *MI, unsigned OpNo, raw_ostream &O);
+  void printUpdateExecMask(const MCInst *MI, unsigned OpNo, raw_ostream &O);
+  void printUpdatePred(const MCInst *MI, unsigned OpNo, raw_ostream &O);
+  void printWrite(const MCInst *MI, unsigned OpNo, raw_ostream &O);
+};
+
+} // End namespace llvm
+
+#endif // AMDGPUINSTRPRINTER_H
diff --git a/lib/Target/R600/InstPrinter/CMakeLists.txt b/lib/Target/R600/InstPrinter/CMakeLists.txt
new file mode 100644
index 0000000000..67763374e1
--- /dev/null
+++ b/lib/Target/R600/InstPrinter/CMakeLists.txt
@@ -0,0 +1,7 @@
+include_directories( ${CMAKE_CURRENT_BINARY_DIR}/.. ${CMAKE_CURRENT_SOURCE_DIR}/.. )
+
+add_llvm_library(LLVMR600AsmPrinter
+  AMDGPUInstPrinter.cpp
+  )
+
+add_dependencies(LLVMR600AsmPrinter R600CommonTableGen)
diff --git a/lib/Target/R600/InstPrinter/LLVMBuild.txt b/lib/Target/R600/InstPrinter/LLVMBuild.txt
new file mode 100644
index 0000000000..ec0be89f10
--- /dev/null
+++ b/lib/Target/R600/InstPrinter/LLVMBuild.txt
@@ -0,0 +1,24 @@
+;===- ./lib/Target/R600/InstPrinter/LLVMBuild.txt -----------*- Conf -*--===;
+;
+;                     The LLVM Compiler Infrastructure
+;
+; This file is distributed under the University of Illinois Open Source
+; License. See LICENSE.TXT for details.
+;
+;===------------------------------------------------------------------------===;
+;
+; This is an LLVMBuild description file for the components in this subdirectory.
+;
+; For more information on the LLVMBuild system, please see:
+;
+;   http://llvm.org/docs/LLVMBuild.html
+;
+;===------------------------------------------------------------------------===;
+
+[component_0]
+type = Library
+name = R600AsmPrinter
+parent = R600
+required_libraries = MC Support
+add_to_library_groups = R600
+
diff --git a/lib/Target/R600/InstPrinter/Makefile b/lib/Target/R600/InstPrinter/Makefile
new file mode 100644
index 0000000000..a794cc1124
--- /dev/null
+++ b/lib/Target/R600/InstPrinter/Makefile
@@ -0,0 +1,15 @@
+#===- lib/Target/R600/AsmPrinter/Makefile ------------------*- Makefile -*-===##
+#
+#                     The LLVM Compiler Infrastructure
+#
+# This file is distributed under the University of Illinois Open Source
+# License. See LICENSE.TXT for details.
+#
+##===----------------------------------------------------------------------===##
+LEVEL = ../../../..
+LIBRARYNAME = LLVMR600AsmPrinter
+
+# Hack: we need to include 'main' x86 target directory to grab private headers
+CPP.Flags += -I$(PROJ_OBJ_DIR)/.. -I$(PROJ_SRC_DIR)/..
+
+include $(LEVEL)/Makefile.common
diff --git a/lib/Target/R600/LLVMBuild.txt b/lib/Target/R600/LLVMBuild.txt
new file mode 100644
index 0000000000..f2a7554e52
--- /dev/null
+++ b/lib/Target/R600/LLVMBuild.txt
@@ -0,0 +1,32 @@
+;===- ./lib/Target/AMDIL/LLVMBuild.txt -------------------------*- Conf -*--===;
+;
+;                     The LLVM Compiler Infrastructure
+;
+; This file is distributed under the University of Illinois Open Source
+; License. See LICENSE.TXT for details.
+;
+;===------------------------------------------------------------------------===;
+;
+; This is an LLVMBuild description file for the components in this subdirectory.
+;
+; For more information on the LLVMBuild system, please see:
+;
+;   http://llvm.org/docs/LLVMBuild.html
+;
+;===------------------------------------------------------------------------===;
+
+[common]
+subdirectories = InstPrinter MCTargetDesc TargetInfo
+
+[component_0]
+type = TargetGroup
+name = R600
+parent = Target
+has_asmprinter = 1
+
+[component_1]
+type = Library
+name = R600CodeGen
+parent = R600
+required_libraries = AsmPrinter CodeGen Core SelectionDAG Support Target MC R600AsmPrinter R600Desc R600Info 
+add_to_library_groups = R600
diff --git a/lib/Target/R600/MCTargetDesc/AMDGPUAsmBackend.cpp b/lib/Target/R600/MCTargetDesc/AMDGPUAsmBackend.cpp
new file mode 100644
index 0000000000..3417fbc1fe
--- /dev/null
+++ b/lib/Target/R600/MCTargetDesc/AMDGPUAsmBackend.cpp
@@ -0,0 +1,82 @@
+//===-- AMDGPUAsmBackend.cpp - AMDGPU Assembler Backend -------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//===----------------------------------------------------------------------===//
+
+#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
+#include "llvm/ADT/StringRef.h"
+#include "llvm/MC/MCAsmBackend.h"
+#include "llvm/MC/MCAssembler.h"
+#include "llvm/MC/MCObjectWriter.h"
+#include "llvm/MC/MCValue.h"
+#include "llvm/Support/TargetRegistry.h"
+
+using namespace llvm;
+
+namespace {
+
+class AMDGPUMCObjectWriter : public MCObjectWriter {
+public:
+  AMDGPUMCObjectWriter(raw_ostream &OS) : MCObjectWriter(OS, true) { }
+  virtual void ExecutePostLayoutBinding(MCAssembler &Asm,
+                                        const MCAsmLayout &Layout) {
+    //XXX: Implement if necessary.
+  }
+  virtual void RecordRelocation(const MCAssembler &Asm,
+                                const MCAsmLayout &Layout,
+                                const MCFragment *Fragment,
+                                const MCFixup &Fixup,
+                                MCValue Target, uint64_t &FixedValue) {
+    assert(!"Not implemented");
+  }
+
+  virtual void WriteObject(MCAssembler &Asm, const MCAsmLayout &Layout);
+
+};
+
+class AMDGPUAsmBackend : public MCAsmBackend {
+public:
+  AMDGPUAsmBackend(const Target &T)
+    : MCAsmBackend() {}
+
+  virtual AMDGPUMCObjectWriter *createObjectWriter(raw_ostream &OS) const;
+  virtual unsigned getNumFixupKinds() const { return 0; };
+  virtual void applyFixup(const MCFixup &Fixup, char *Data, unsigned DataSize,
+                          uint64_t Value) const { assert(!"Not implemented"); }
+  virtual bool fixupNeedsRelaxation(const MCFixup &Fixup, uint64_t Value,
+                                    const MCInstFragment *DF,
+                                    const MCAsmLayout &Layout) const {
+    return false;
+  }
+  virtual void relaxInstruction(const MCInst &Inst, MCInst &Res) const {
+    assert(!"Not implemented");
+  }
+  virtual bool mayNeedRelaxation(const MCInst &Inst) const { return false; }
+  virtual bool writeNopData(uint64_t Count, MCObjectWriter *OW) const {
+    return true;
+  }
+};
+
+} //End anonymous namespace
+
+void AMDGPUMCObjectWriter::WriteObject(MCAssembler &Asm,
+                                       const MCAsmLayout &Layout) {
+  for (MCAssembler::iterator I = Asm.begin(), E = Asm.end(); I != E; ++I) {
+    Asm.writeSectionData(I, Layout);
+  }
+}
+
+MCAsmBackend *llvm::createAMDGPUAsmBackend(const Target &T, StringRef TT,
+                                           StringRef CPU) {
+  return new AMDGPUAsmBackend(T);
+}
+
+AMDGPUMCObjectWriter * AMDGPUAsmBackend::createObjectWriter(
+                                                        raw_ostream &OS) const {
+  return new AMDGPUMCObjectWriter(OS);
+}
diff --git a/lib/Target/R600/MCTargetDesc/AMDGPUMCAsmInfo.cpp b/lib/Target/R600/MCTargetDesc/AMDGPUMCAsmInfo.cpp
new file mode 100644
index 0000000000..4d3d3e7945
--- /dev/null
+++ b/lib/Target/R600/MCTargetDesc/AMDGPUMCAsmInfo.cpp
@@ -0,0 +1,85 @@
+//===-- MCTargetDesc/AMDGPUMCAsmInfo.cpp - Assembly Info ------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUMCAsmInfo.h"
+
+using namespace llvm;
+AMDGPUMCAsmInfo::AMDGPUMCAsmInfo(const Target &T, StringRef &TT) : MCAsmInfo() {
+  HasSingleParameterDotFile = false;
+  WeakDefDirective = 0;
+  //===------------------------------------------------------------------===//
+  HasSubsectionsViaSymbols = true;
+  HasMachoZeroFillDirective = false;
+  HasMachoTBSSDirective = false;
+  HasStaticCtorDtorReferenceInStaticMode = false;
+  LinkerRequiresNonEmptyDwarfLines = true;
+  MaxInstLength = 16;
+  PCSymbol = "$";
+  SeparatorString = "\n";
+  CommentColumn = 40;
+  CommentString = ";";
+  LabelSuffix = ":";
+  GlobalPrefix = "@";
+  PrivateGlobalPrefix = ";.";
+  LinkerPrivateGlobalPrefix = "!";
+  InlineAsmStart = ";#ASMSTART";
+  InlineAsmEnd = ";#ASMEND";
+  AssemblerDialect = 0;
+  AllowQuotesInName = false;
+  AllowNameToStartWithDigit = false;
+  AllowPeriodsInName = false;
+
+  //===--- Data Emission Directives -------------------------------------===//
+  ZeroDirective = ".zero";
+  AsciiDirective = ".ascii\t";
+  AscizDirective = ".asciz\t";
+  Data8bitsDirective = ".byte\t";
+  Data16bitsDirective = ".short\t";
+  Data32bitsDirective = ".long\t";
+  Data64bitsDirective = ".quad\t";
+  GPRel32Directive = 0;
+  SunStyleELFSectionSwitchSyntax = true;
+  UsesELFSectionDirectiveForBSS = true;
+  HasMicrosoftFastStdCallMangling = false;
+
+  //===--- Alignment Information ----------------------------------------===//
+  AlignDirective = ".align\t";
+  AlignmentIsInBytes = true;
+  TextAlignFillValue = 0;
+
+  //===--- Global Variable Emission Directives --------------------------===//
+  GlobalDirective = ".global";
+  ExternDirective = ".extern";
+  HasSetDirective = false;
+  HasAggressiveSymbolFolding = true;
+  COMMDirectiveAlignmentIsInBytes = false;
+  HasDotTypeDotSizeDirective = false;
+  HasNoDeadStrip = true;
+  HasSymbolResolver = false;
+  WeakRefDirective = ".weakref\t";
+  LinkOnceDirective = 0;
+  //===--- Dwarf Emission Directives -----------------------------------===//
+  HasLEB128 = true;
+  SupportsDebugInformation = true;
+  ExceptionsType = ExceptionHandling::None;
+  DwarfUsesInlineInfoSection = false;
+  DwarfSectionOffsetDirective = ".offset";
+
+}
+
+const char*
+AMDGPUMCAsmInfo::getDataASDirective(unsigned int Size, unsigned int AS) const {
+  return 0;
+}
+
+const MCSection*
+AMDGPUMCAsmInfo::getNonexecutableStackSection(MCContext &CTX) const {
+  return 0;
+}
diff --git a/lib/Target/R600/MCTargetDesc/AMDGPUMCAsmInfo.h b/lib/Target/R600/MCTargetDesc/AMDGPUMCAsmInfo.h
new file mode 100644
index 0000000000..3ad0fa6824
--- /dev/null
+++ b/lib/Target/R600/MCTargetDesc/AMDGPUMCAsmInfo.h
@@ -0,0 +1,30 @@
+//===-- MCTargetDesc/AMDGPUMCAsmInfo.h - AMDGPU MCAsm Interface  ----------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef AMDGPUMCASMINFO_H
+#define AMDGPUMCASMINFO_H
+
+#include "llvm/MC/MCAsmInfo.h"
+namespace llvm {
+
+class Target;
+class StringRef;
+
+class AMDGPUMCAsmInfo : public MCAsmInfo {
+public:
+  explicit AMDGPUMCAsmInfo(const Target &T, StringRef &TT);
+  const char* getDataASDirective(unsigned int Size, unsigned int AS) const;
+  const MCSection* getNonexecutableStackSection(MCContext &CTX) const;
+};
+} // namespace llvm
+#endif // AMDGPUMCASMINFO_H
diff --git a/lib/Target/R600/MCTargetDesc/AMDGPUMCCodeEmitter.h b/lib/Target/R600/MCTargetDesc/AMDGPUMCCodeEmitter.h
new file mode 100644
index 0000000000..9d0d6cf6fd
--- /dev/null
+++ b/lib/Target/R600/MCTargetDesc/AMDGPUMCCodeEmitter.h
@@ -0,0 +1,60 @@
+//===-- AMDGPUCodeEmitter.h - AMDGPU Code Emitter interface -----------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief CodeEmitter interface for R600 and SI codegen.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef AMDGPUCODEEMITTER_H
+#define AMDGPUCODEEMITTER_H
+
+#include "llvm/MC/MCCodeEmitter.h"
+#include "llvm/Support/raw_ostream.h"
+
+namespace llvm {
+
+class MCInst;
+class MCOperand;
+
+class AMDGPUMCCodeEmitter : public MCCodeEmitter {
+public:
+
+  uint64_t getBinaryCodeForInstr(const MCInst &MI,
+                                 SmallVectorImpl<MCFixup> &Fixups) const;
+
+  virtual uint64_t getMachineOpValue(const MCInst &MI, const MCOperand &MO,
+                                     SmallVectorImpl<MCFixup> &Fixups) const {
+    return 0;
+  }
+
+  virtual unsigned GPR4AlignEncode(const MCInst  &MI, unsigned OpNo,
+                                   SmallVectorImpl<MCFixup> &Fixups) const {
+    return 0;
+  }
+  virtual unsigned GPR2AlignEncode(const MCInst &MI, unsigned OpNo,
+                                   SmallVectorImpl<MCFixup> &Fixups) const {
+    return 0;
+  }
+  virtual uint64_t VOPPostEncode(const MCInst &MI, uint64_t Value) const {
+    return Value;
+  }
+  virtual uint64_t i32LiteralEncode(const MCInst &MI, unsigned OpNo,
+                                   SmallVectorImpl<MCFixup> &Fixups) const {
+    return 0;
+  }
+  virtual uint32_t SMRDmemriEncode(const MCInst &MI, unsigned OpNo,
+                                   SmallVectorImpl<MCFixup> &Fixups) const {
+    return 0;
+  }
+};
+
+} // End namespace llvm
+
+#endif // AMDGPUCODEEMITTER_H
diff --git a/lib/Target/R600/MCTargetDesc/AMDGPUMCTargetDesc.cpp b/lib/Target/R600/MCTargetDesc/AMDGPUMCTargetDesc.cpp
new file mode 100644
index 0000000000..6a62856e10
--- /dev/null
+++ b/lib/Target/R600/MCTargetDesc/AMDGPUMCTargetDesc.cpp
@@ -0,0 +1,113 @@
+//===-- AMDGPUMCTargetDesc.cpp - AMDGPU Target Descriptions ---------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief This file provides AMDGPU specific target descriptions.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPUMCTargetDesc.h"
+#include "AMDGPUMCAsmInfo.h"
+#include "InstPrinter/AMDGPUInstPrinter.h"
+#include "llvm/MC/MachineLocation.h"
+#include "llvm/MC/MCCodeGenInfo.h"
+#include "llvm/MC/MCInstrInfo.h"
+#include "llvm/MC/MCRegisterInfo.h"
+#include "llvm/MC/MCStreamer.h"
+#include "llvm/MC/MCSubtargetInfo.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/TargetRegistry.h"
+
+#define GET_INSTRINFO_MC_DESC
+#include "AMDGPUGenInstrInfo.inc"
+
+#define GET_SUBTARGETINFO_MC_DESC
+#include "AMDGPUGenSubtargetInfo.inc"
+
+#define GET_REGINFO_MC_DESC
+#include "AMDGPUGenRegisterInfo.inc"
+
+using namespace llvm;
+
+static MCInstrInfo *createAMDGPUMCInstrInfo() {
+  MCInstrInfo *X = new MCInstrInfo();
+  InitAMDGPUMCInstrInfo(X);
+  return X;
+}
+
+static MCRegisterInfo *createAMDGPUMCRegisterInfo(StringRef TT) {
+  MCRegisterInfo *X = new MCRegisterInfo();
+  InitAMDGPUMCRegisterInfo(X, 0);
+  return X;
+}
+
+static MCSubtargetInfo *createAMDGPUMCSubtargetInfo(StringRef TT, StringRef CPU,
+                                                   StringRef FS) {
+  MCSubtargetInfo * X = new MCSubtargetInfo();
+  InitAMDGPUMCSubtargetInfo(X, TT, CPU, FS);
+  return X;
+}
+
+static MCCodeGenInfo *createAMDGPUMCCodeGenInfo(StringRef TT, Reloc::Model RM,
+                                               CodeModel::Model CM,
+                                               CodeGenOpt::Level OL) {
+  MCCodeGenInfo *X = new MCCodeGenInfo();
+  X->InitMCCodeGenInfo(RM, CM, OL);
+  return X;
+}
+
+static MCInstPrinter *createAMDGPUMCInstPrinter(const Target &T,
+                                                unsigned SyntaxVariant,
+                                                const MCAsmInfo &MAI,
+                                                const MCInstrInfo &MII,
+                                                const MCRegisterInfo &MRI,
+                                                const MCSubtargetInfo &STI) {
+  return new AMDGPUInstPrinter(MAI, MII, MRI);
+}
+
+static MCCodeEmitter *createAMDGPUMCCodeEmitter(const MCInstrInfo &MCII,
+                                                const MCRegisterInfo &MRI,
+                                                const MCSubtargetInfo &STI,
+                                                MCContext &Ctx) {
+  if (STI.getFeatureBits() & AMDGPU::Feature64BitPtr) {
+    return createSIMCCodeEmitter(MCII, MRI, STI, Ctx);
+  } else {
+    return createR600MCCodeEmitter(MCII, MRI, STI, Ctx);
+  }
+}
+
+static MCStreamer *createMCStreamer(const Target &T, StringRef TT,
+                                    MCContext &Ctx, MCAsmBackend &MAB,
+                                    raw_ostream &_OS,
+                                    MCCodeEmitter *_Emitter,
+                                    bool RelaxAll,
+                                    bool NoExecStack) {
+  return createPureStreamer(Ctx, MAB, _OS, _Emitter);
+}
+
+extern "C" void LLVMInitializeR600TargetMC() {
+
+  RegisterMCAsmInfo<AMDGPUMCAsmInfo> Y(TheAMDGPUTarget);
+
+  TargetRegistry::RegisterMCCodeGenInfo(TheAMDGPUTarget, createAMDGPUMCCodeGenInfo);
+
+  TargetRegistry::RegisterMCInstrInfo(TheAMDGPUTarget, createAMDGPUMCInstrInfo);
+
+  TargetRegistry::RegisterMCRegInfo(TheAMDGPUTarget, createAMDGPUMCRegisterInfo);
+
+  TargetRegistry::RegisterMCSubtargetInfo(TheAMDGPUTarget, createAMDGPUMCSubtargetInfo);
+
+  TargetRegistry::RegisterMCInstPrinter(TheAMDGPUTarget, createAMDGPUMCInstPrinter);
+
+  TargetRegistry::RegisterMCCodeEmitter(TheAMDGPUTarget, createAMDGPUMCCodeEmitter);
+
+  TargetRegistry::RegisterMCAsmBackend(TheAMDGPUTarget, createAMDGPUAsmBackend);
+
+  TargetRegistry::RegisterMCObjectStreamer(TheAMDGPUTarget, createMCStreamer);
+}
diff --git a/lib/Target/R600/MCTargetDesc/AMDGPUMCTargetDesc.h b/lib/Target/R600/MCTargetDesc/AMDGPUMCTargetDesc.h
new file mode 100644
index 0000000000..363a4af3f3
--- /dev/null
+++ b/lib/Target/R600/MCTargetDesc/AMDGPUMCTargetDesc.h
@@ -0,0 +1,55 @@
+//===-- AMDGPUMCTargetDesc.h - AMDGPU Target Descriptions -----*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Provides AMDGPU specific target descriptions.
+//
+//===----------------------------------------------------------------------===//
+//
+
+#ifndef AMDGPUMCTARGETDESC_H
+#define AMDGPUMCTARGETDESC_H
+
+#include "llvm/ADT/StringRef.h"
+
+namespace llvm {
+class MCAsmBackend;
+class MCCodeEmitter;
+class MCContext;
+class MCInstrInfo;
+class MCRegisterInfo;
+class MCSubtargetInfo;
+class Target;
+
+extern Target TheAMDGPUTarget;
+
+MCCodeEmitter *createR600MCCodeEmitter(const MCInstrInfo &MCII,
+                                       const MCRegisterInfo &MRI,
+                                       const MCSubtargetInfo &STI,
+                                       MCContext &Ctx);
+
+MCCodeEmitter *createSIMCCodeEmitter(const MCInstrInfo &MCII,
+                                     const MCRegisterInfo &MRI,
+                                     const MCSubtargetInfo &STI,
+                                     MCContext &Ctx);
+
+MCAsmBackend *createAMDGPUAsmBackend(const Target &T, StringRef TT,
+                                     StringRef CPU);
+} // End llvm namespace
+
+#define GET_REGINFO_ENUM
+#include "AMDGPUGenRegisterInfo.inc"
+
+#define GET_INSTRINFO_ENUM
+#include "AMDGPUGenInstrInfo.inc"
+
+#define GET_SUBTARGETINFO_ENUM
+#include "AMDGPUGenSubtargetInfo.inc"
+
+#endif // AMDGPUMCTARGETDESC_H
diff --git a/lib/Target/R600/MCTargetDesc/CMakeLists.txt b/lib/Target/R600/MCTargetDesc/CMakeLists.txt
new file mode 100644
index 0000000000..37e714c2e7
--- /dev/null
+++ b/lib/Target/R600/MCTargetDesc/CMakeLists.txt
@@ -0,0 +1,10 @@
+
+add_llvm_library(LLVMR600Desc
+  AMDGPUAsmBackend.cpp
+  AMDGPUMCTargetDesc.cpp
+  AMDGPUMCAsmInfo.cpp
+  R600MCCodeEmitter.cpp
+  SIMCCodeEmitter.cpp
+  )
+
+add_dependencies(LLVMR600Desc AMDGPUCommonTableGen)
diff --git a/lib/Target/R600/MCTargetDesc/LLVMBuild.txt b/lib/Target/R600/MCTargetDesc/LLVMBuild.txt
new file mode 100644
index 0000000000..b1beab0bb3
--- /dev/null
+++ b/lib/Target/R600/MCTargetDesc/LLVMBuild.txt
@@ -0,0 +1,23 @@
+;===- ./lib/Target/R600/MCTargetDesc/LLVMBuild.txt ------------*- Conf -*--===;
+;
+;                     The LLVM Compiler Infrastructure
+;
+; This file is distributed under the University of Illinois Open Source
+; License. See LICENSE.TXT for details.
+;
+;===------------------------------------------------------------------------===;
+;
+; This is an LLVMBuild description file for the components in this subdirectory.
+;
+; For more information on the LLVMBuild system, please see:
+;
+;   http://llvm.org/docs/LLVMBuild.html
+;
+;===------------------------------------------------------------------------===;
+
+[component_0]
+type = Library
+name = R600Desc
+parent = R600
+required_libraries = R600AsmPrinter R600Info MC
+add_to_library_groups = R600
diff --git a/lib/Target/R600/MCTargetDesc/Makefile b/lib/Target/R600/MCTargetDesc/Makefile
new file mode 100644
index 0000000000..8894a7607f
--- /dev/null
+++ b/lib/Target/R600/MCTargetDesc/Makefile
@@ -0,0 +1,16 @@
+##===- lib/Target/AMDGPU/TargetDesc/Makefile ----------------*- Makefile -*-===##
+#
+#                     The LLVM Compiler Infrastructure
+#
+# This file is distributed under the University of Illinois Open Source
+# License. See LICENSE.TXT for details.
+#
+##===----------------------------------------------------------------------===##
+
+LEVEL = ../../../..
+LIBRARYNAME = LLVMR600Desc
+
+# Hack: we need to include 'main' target directory to grab private headers
+CPP.Flags += -I$(PROJ_OBJ_DIR)/.. -I$(PROJ_SRC_DIR)/..
+
+include $(LEVEL)/Makefile.common
diff --git a/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp b/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
new file mode 100644
index 0000000000..dc91924c73
--- /dev/null
+++ b/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
@@ -0,0 +1,575 @@
+//===- R600MCCodeEmitter.cpp - Code Emitter for R600->Cayman GPU families -===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+///
+/// This code emitter outputs bytecode that is understood by the r600g driver
+/// in the Mesa [1] project.  The bytecode is very similar to the hardware's ISA,
+/// but it still needs to be run through a finalizer in order to be executed
+/// by the GPU.
+///
+/// [1] http://www.mesa3d.org/
+//
+//===----------------------------------------------------------------------===//
+
+#include "R600Defines.h"
+#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
+#include "MCTargetDesc/AMDGPUMCCodeEmitter.h"
+#include "llvm/MC/MCCodeEmitter.h"
+#include "llvm/MC/MCContext.h"
+#include "llvm/MC/MCInst.h"
+#include "llvm/MC/MCInstrInfo.h"
+#include "llvm/MC/MCRegisterInfo.h"
+#include "llvm/MC/MCSubtargetInfo.h"
+#include "llvm/Support/raw_ostream.h"
+
+#include <stdio.h>
+
+#define SRC_BYTE_COUNT 11
+#define DST_BYTE_COUNT 5
+
+using namespace llvm;
+
+namespace {
+
+class R600MCCodeEmitter : public AMDGPUMCCodeEmitter {
+  R600MCCodeEmitter(const R600MCCodeEmitter &); // DO NOT IMPLEMENT
+  void operator=(const R600MCCodeEmitter &); // DO NOT IMPLEMENT
+  const MCInstrInfo &MCII;
+  const MCRegisterInfo &MRI;
+  const MCSubtargetInfo &STI;
+  MCContext &Ctx;
+
+public:
+
+  R600MCCodeEmitter(const MCInstrInfo &mcii, const MCRegisterInfo &mri,
+                    const MCSubtargetInfo &sti, MCContext &ctx)
+    : MCII(mcii), MRI(mri), STI(sti), Ctx(ctx) { }
+
+  /// \brief Encode the instruction and write it to the OS.
+  virtual void EncodeInstruction(const MCInst &MI, raw_ostream &OS,
+                         SmallVectorImpl<MCFixup> &Fixups) const;
+
+  /// \returns the encoding for an MCOperand.
+  virtual uint64_t getMachineOpValue(const MCInst &MI, const MCOperand &MO,
+                                     SmallVectorImpl<MCFixup> &Fixups) const;
+private:
+
+  void EmitALUInstr(const MCInst &MI, SmallVectorImpl<MCFixup> &Fixups,
+                    raw_ostream &OS) const;
+  void EmitSrc(const MCInst &MI, unsigned OpIdx, raw_ostream &OS) const;
+  void EmitSrcISA(const MCInst &MI, unsigned OpIdx, uint64_t &Value,
+                  raw_ostream &OS) const;
+  void EmitDst(const MCInst &MI, raw_ostream &OS) const;
+  void EmitTexInstr(const MCInst &MI, SmallVectorImpl<MCFixup> &Fixups,
+                    raw_ostream &OS) const;
+  void EmitFCInstr(const MCInst &MI, raw_ostream &OS) const;
+
+  void EmitNullBytes(unsigned int byteCount, raw_ostream &OS) const;
+
+  void EmitByte(unsigned int byte, raw_ostream &OS) const;
+
+  void EmitTwoBytes(uint32_t bytes, raw_ostream &OS) const;
+
+  void Emit(uint32_t value, raw_ostream &OS) const;
+  void Emit(uint64_t value, raw_ostream &OS) const;
+
+  unsigned getHWRegChan(unsigned reg) const;
+  unsigned getHWReg(unsigned regNo) const;
+
+  bool isFCOp(unsigned opcode) const;
+  bool isTexOp(unsigned opcode) const;
+  bool isFlagSet(const MCInst &MI, unsigned Operand, unsigned Flag) const;
+
+};
+
+} // End anonymous namespace
+
+enum RegElement {
+  ELEMENT_X = 0,
+  ELEMENT_Y,
+  ELEMENT_Z,
+  ELEMENT_W
+};
+
+enum InstrTypes {
+  INSTR_ALU = 0,
+  INSTR_TEX,
+  INSTR_FC,
+  INSTR_NATIVE,
+  INSTR_VTX,
+  INSTR_EXPORT
+};
+
+enum FCInstr {
+  FC_IF_PREDICATE = 0,
+  FC_ELSE,
+  FC_ENDIF,
+  FC_BGNLOOP,
+  FC_ENDLOOP,
+  FC_BREAK_PREDICATE,
+  FC_CONTINUE
+};
+
+enum TextureTypes {
+  TEXTURE_1D = 1,
+  TEXTURE_2D,
+  TEXTURE_3D,
+  TEXTURE_CUBE,
+  TEXTURE_RECT,
+  TEXTURE_SHADOW1D,
+  TEXTURE_SHADOW2D,
+  TEXTURE_SHADOWRECT,
+  TEXTURE_1D_ARRAY,
+  TEXTURE_2D_ARRAY,
+  TEXTURE_SHADOW1D_ARRAY,
+  TEXTURE_SHADOW2D_ARRAY
+};
+
+MCCodeEmitter *llvm::createR600MCCodeEmitter(const MCInstrInfo &MCII,
+                                           const MCRegisterInfo &MRI,
+                                           const MCSubtargetInfo &STI,
+                                           MCContext &Ctx) {
+  return new R600MCCodeEmitter(MCII, MRI, STI, Ctx);
+}
+
+void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
+                                       SmallVectorImpl<MCFixup> &Fixups) const {
+  if (isTexOp(MI.getOpcode())) {
+    EmitTexInstr(MI, Fixups, OS);
+  } else if (isFCOp(MI.getOpcode())){
+    EmitFCInstr(MI, OS);
+  } else if (MI.getOpcode() == AMDGPU::RETURN ||
+    MI.getOpcode() == AMDGPU::BUNDLE ||
+    MI.getOpcode() == AMDGPU::KILL) {
+    return;
+  } else {
+    switch(MI.getOpcode()) {
+    case AMDGPU::RAT_WRITE_CACHELESS_32_eg:
+    case AMDGPU::RAT_WRITE_CACHELESS_128_eg: {
+      uint64_t inst = getBinaryCodeForInstr(MI, Fixups);
+      EmitByte(INSTR_NATIVE, OS);
+      Emit(inst, OS);
+      break;
+    }
+    case AMDGPU::CONSTANT_LOAD_eg:
+    case AMDGPU::VTX_READ_PARAM_8_eg:
+    case AMDGPU::VTX_READ_PARAM_16_eg:
+    case AMDGPU::VTX_READ_PARAM_32_eg:
+    case AMDGPU::VTX_READ_GLOBAL_8_eg:
+    case AMDGPU::VTX_READ_GLOBAL_32_eg:
+    case AMDGPU::VTX_READ_GLOBAL_128_eg: {
+      uint64_t InstWord01 = getBinaryCodeForInstr(MI, Fixups);
+      uint32_t InstWord2 = MI.getOperand(2).getImm(); // Offset
+
+      EmitByte(INSTR_VTX, OS);
+      Emit(InstWord01, OS);
+      Emit(InstWord2, OS);
+      break;
+    }
+    case AMDGPU::EG_ExportSwz:
+    case AMDGPU::R600_ExportSwz:
+    case AMDGPU::EG_ExportBuf:
+    case AMDGPU::R600_ExportBuf: {
+      uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
+      EmitByte(INSTR_EXPORT, OS);
+      Emit(Inst, OS);
+      break;
+    }
+
+    default:
+      EmitALUInstr(MI, Fixups, OS);
+      break;
+    }
+  }
+}
+
+void R600MCCodeEmitter::EmitALUInstr(const MCInst &MI,
+                                     SmallVectorImpl<MCFixup> &Fixups,
+                                     raw_ostream &OS) const {
+  const MCInstrDesc &MCDesc = MCII.get(MI.getOpcode());
+  unsigned NumOperands = MI.getNumOperands();
+
+  // Emit instruction type
+  EmitByte(INSTR_ALU, OS);
+
+  uint64_t InstWord01 = getBinaryCodeForInstr(MI, Fixups);
+
+  //older alu have different encoding for instructions with one or two src
+  //parameters.
+  if ((STI.getFeatureBits() & AMDGPU::FeatureR600ALUInst) &&
+      !(MCDesc.TSFlags & R600_InstFlag::OP3)) {
+    uint64_t ISAOpCode = InstWord01 & (0x3FFULL << 39);
+    InstWord01 &= ~(0x3FFULL << 39);
+    InstWord01 |= ISAOpCode << 1;
+  }
+
+  unsigned SrcIdx = 0;
+  for (unsigned int OpIdx = 1; OpIdx < NumOperands; ++OpIdx) {
+    if (MI.getOperand(OpIdx).isImm() || MI.getOperand(OpIdx).isFPImm() ||
+        OpIdx == (unsigned)MCDesc.findFirstPredOperandIdx()) {
+      continue;
+    }
+    EmitSrcISA(MI, OpIdx, InstWord01, OS);
+    SrcIdx++;
+  }
+
+  // Emit zeros for unused sources
+  for ( ; SrcIdx < 3; SrcIdx++) {
+    EmitNullBytes(SRC_BYTE_COUNT - 6, OS);
+  }
+
+  Emit(InstWord01, OS);
+  return;
+}
+
+void R600MCCodeEmitter::EmitSrc(const MCInst &MI, unsigned OpIdx,
+                                raw_ostream &OS) const {
+  const MCOperand &MO = MI.getOperand(OpIdx);
+  union {
+    float f;
+    uint32_t i;
+  } Value;
+  Value.i = 0;
+  // Emit the source select (2 bytes).  For GPRs, this is the register index.
+  // For other potential instruction operands, (e.g. constant registers) the
+  // value of the source select is defined in the r600isa docs.
+  if (MO.isReg()) {
+    unsigned reg = MO.getReg();
+    EmitTwoBytes(getHWReg(reg), OS);
+    if (reg == AMDGPU::ALU_LITERAL_X) {
+      unsigned ImmOpIndex = MI.getNumOperands() - 1;
+      MCOperand ImmOp = MI.getOperand(ImmOpIndex);
+      if (ImmOp.isFPImm()) {
+        Value.f = ImmOp.getFPImm();
+      } else {
+        assert(ImmOp.isImm());
+        Value.i = ImmOp.getImm();
+      }
+    }
+  } else {
+    // XXX: Handle other operand types.
+    EmitTwoBytes(0, OS);
+  }
+
+  // Emit the source channel (1 byte)
+  if (MO.isReg()) {
+    EmitByte(getHWRegChan(MO.getReg()), OS);
+  } else {
+    EmitByte(0, OS);
+  }
+
+  // XXX: Emit isNegated (1 byte)
+  if ((!(isFlagSet(MI, OpIdx, MO_FLAG_ABS)))
+      && (isFlagSet(MI, OpIdx, MO_FLAG_NEG) ||
+     (MO.isReg() &&
+      (MO.getReg() == AMDGPU::NEG_ONE || MO.getReg() == AMDGPU::NEG_HALF)))){
+    EmitByte(1, OS);
+  } else {
+    EmitByte(0, OS);
+  }
+
+  // Emit isAbsolute (1 byte)
+  if (isFlagSet(MI, OpIdx, MO_FLAG_ABS)) {
+    EmitByte(1, OS);
+  } else {
+    EmitByte(0, OS);
+  }
+
+  // XXX: Emit relative addressing mode (1 byte)
+  EmitByte(0, OS);
+
+  // Emit kc_bank, This will be adjusted later by r600_asm
+  EmitByte(0, OS);
+
+  // Emit the literal value, if applicable (4 bytes).
+  Emit(Value.i, OS);
+
+}
+
+void R600MCCodeEmitter::EmitSrcISA(const MCInst &MI, unsigned OpIdx,
+                                   uint64_t &Value, raw_ostream &OS) const {
+  const MCOperand &MO = MI.getOperand(OpIdx);
+  union {
+    float f;
+    uint32_t i;
+  } InlineConstant;
+  InlineConstant.i = 0;
+  // Emit the source select (2 bytes).  For GPRs, this is the register index.
+  // For other potential instruction operands, (e.g. constant registers) the
+  // value of the source select is defined in the r600isa docs.
+  if (MO.isReg()) {
+    unsigned Reg = MO.getReg();
+    if (AMDGPUMCRegisterClasses[AMDGPU::R600_CReg32RegClassID].contains(Reg)) {
+      EmitByte(1, OS);
+    } else {
+      EmitByte(0, OS);
+    }
+
+    if (Reg == AMDGPU::ALU_LITERAL_X) {
+      unsigned ImmOpIndex = MI.getNumOperands() - 1;
+      MCOperand ImmOp = MI.getOperand(ImmOpIndex);
+      if (ImmOp.isFPImm()) {
+        InlineConstant.f = ImmOp.getFPImm();
+      } else {
+        assert(ImmOp.isImm());
+        InlineConstant.i = ImmOp.getImm();
+      }
+    }
+  }
+
+  // Emit the literal value, if applicable (4 bytes).
+  Emit(InlineConstant.i, OS);
+}
+
+void R600MCCodeEmitter::EmitTexInstr(const MCInst &MI,
+                                     SmallVectorImpl<MCFixup> &Fixups,
+                                     raw_ostream &OS) const {
+
+  unsigned Opcode = MI.getOpcode();
+  bool hasOffsets = (Opcode == AMDGPU::TEX_LD);
+  unsigned OpOffset = hasOffsets ? 3 : 0;
+  int64_t Resource = MI.getOperand(OpOffset + 2).getImm();
+  int64_t Sampler = MI.getOperand(OpOffset + 3).getImm();
+  int64_t TextureType = MI.getOperand(OpOffset + 4).getImm();
+  unsigned srcSelect[4] = {0, 1, 2, 3};
+
+  // Emit instruction type
+  EmitByte(1, OS);
+
+  // Emit instruction
+  EmitByte(getBinaryCodeForInstr(MI, Fixups), OS);
+
+  // Emit resource id
+  EmitByte(Resource, OS);
+
+  // Emit source register
+  EmitByte(getHWReg(MI.getOperand(1).getReg()), OS);
+
+  // XXX: Emit src isRelativeAddress
+  EmitByte(0, OS);
+
+  // Emit destination register
+  EmitByte(getHWReg(MI.getOperand(0).getReg()), OS);
+
+  // XXX: Emit dst isRealtiveAddress
+  EmitByte(0, OS);
+
+  // XXX: Emit dst select
+  EmitByte(0, OS); // X
+  EmitByte(1, OS); // Y
+  EmitByte(2, OS); // Z
+  EmitByte(3, OS); // W
+
+  // XXX: Emit lod bias
+  EmitByte(0, OS);
+
+  // XXX: Emit coord types
+  unsigned coordType[4] = {1, 1, 1, 1};
+
+  if (TextureType == TEXTURE_RECT
+      || TextureType == TEXTURE_SHADOWRECT) {
+    coordType[ELEMENT_X] = 0;
+    coordType[ELEMENT_Y] = 0;
+  }
+
+  if (TextureType == TEXTURE_1D_ARRAY
+      || TextureType == TEXTURE_SHADOW1D_ARRAY) {
+    if (Opcode == AMDGPU::TEX_SAMPLE_C_L || Opcode == AMDGPU::TEX_SAMPLE_C_LB) {
+      coordType[ELEMENT_Y] = 0;
+    } else {
+      coordType[ELEMENT_Z] = 0;
+      srcSelect[ELEMENT_Z] = ELEMENT_Y;
+    }
+  } else if (TextureType == TEXTURE_2D_ARRAY
+             || TextureType == TEXTURE_SHADOW2D_ARRAY) {
+    coordType[ELEMENT_Z] = 0;
+  }
+
+  for (unsigned i = 0; i < 4; i++) {
+    EmitByte(coordType[i], OS);
+  }
+
+  // XXX: Emit offsets
+  if (hasOffsets)
+	  for (unsigned i = 2; i < 5; i++)
+		  EmitByte(MI.getOperand(i).getImm()<<1, OS);
+  else
+	  EmitNullBytes(3, OS);
+
+  // Emit sampler id
+  EmitByte(Sampler, OS);
+
+  // XXX:Emit source select
+  if ((TextureType == TEXTURE_SHADOW1D
+      || TextureType == TEXTURE_SHADOW2D
+      || TextureType == TEXTURE_SHADOWRECT
+      || TextureType == TEXTURE_SHADOW1D_ARRAY)
+      && Opcode != AMDGPU::TEX_SAMPLE_C_L
+      && Opcode != AMDGPU::TEX_SAMPLE_C_LB) {
+    srcSelect[ELEMENT_W] = ELEMENT_Z;
+  }
+
+  for (unsigned i = 0; i < 4; i++) {
+    EmitByte(srcSelect[i], OS);
+  }
+}
+
+void R600MCCodeEmitter::EmitFCInstr(const MCInst &MI, raw_ostream &OS) const {
+
+  // Emit instruction type
+  EmitByte(INSTR_FC, OS);
+
+  // Emit SRC
+  unsigned NumOperands = MI.getNumOperands();
+  if (NumOperands > 0) {
+    assert(NumOperands == 1);
+    EmitSrc(MI, 0, OS);
+  } else {
+    EmitNullBytes(SRC_BYTE_COUNT, OS);
+  }
+
+  // Emit FC Instruction
+  enum FCInstr instr;
+  switch (MI.getOpcode()) {
+  case AMDGPU::PREDICATED_BREAK:
+    instr = FC_BREAK_PREDICATE;
+    break;
+  case AMDGPU::CONTINUE:
+    instr = FC_CONTINUE;
+    break;
+  case AMDGPU::IF_PREDICATE_SET:
+    instr = FC_IF_PREDICATE;
+    break;
+  case AMDGPU::ELSE:
+    instr = FC_ELSE;
+    break;
+  case AMDGPU::ENDIF:
+    instr = FC_ENDIF;
+    break;
+  case AMDGPU::ENDLOOP:
+    instr = FC_ENDLOOP;
+    break;
+  case AMDGPU::WHILELOOP:
+    instr = FC_BGNLOOP;
+    break;
+  default:
+    abort();
+    break;
+  }
+  EmitByte(instr, OS);
+}
+
+void R600MCCodeEmitter::EmitNullBytes(unsigned int ByteCount,
+                                      raw_ostream &OS) const {
+
+  for (unsigned int i = 0; i < ByteCount; i++) {
+    EmitByte(0, OS);
+  }
+}
+
+void R600MCCodeEmitter::EmitByte(unsigned int Byte, raw_ostream &OS) const {
+  OS.write((uint8_t) Byte & 0xff);
+}
+
+void R600MCCodeEmitter::EmitTwoBytes(unsigned int Bytes,
+                                     raw_ostream &OS) const {
+  OS.write((uint8_t) (Bytes & 0xff));
+  OS.write((uint8_t) ((Bytes >> 8) & 0xff));
+}
+
+void R600MCCodeEmitter::Emit(uint32_t Value, raw_ostream &OS) const {
+  for (unsigned i = 0; i < 4; i++) {
+    OS.write((uint8_t) ((Value >> (8 * i)) & 0xff));
+  }
+}
+
+void R600MCCodeEmitter::Emit(uint64_t Value, raw_ostream &OS) const {
+  for (unsigned i = 0; i < 8; i++) {
+    EmitByte((Value >> (8 * i)) & 0xff, OS);
+  }
+}
+
+unsigned R600MCCodeEmitter::getHWRegChan(unsigned reg) const {
+  return MRI.getEncodingValue(reg) >> HW_CHAN_SHIFT;
+}
+
+unsigned R600MCCodeEmitter::getHWReg(unsigned RegNo) const {
+  return MRI.getEncodingValue(RegNo) & HW_REG_MASK;
+}
+
+uint64_t R600MCCodeEmitter::getMachineOpValue(const MCInst &MI,
+                                              const MCOperand &MO,
+                                        SmallVectorImpl<MCFixup> &Fixup) const {
+  if (MO.isReg()) {
+    if (HAS_NATIVE_OPERANDS(MCII.get(MI.getOpcode()).TSFlags)) {
+      return MRI.getEncodingValue(MO.getReg());
+    } else {
+      return getHWReg(MO.getReg());
+    }
+  } else if (MO.isImm()) {
+    return MO.getImm();
+  } else {
+    assert(0);
+    return 0;
+  }
+}
+
+//===----------------------------------------------------------------------===//
+// Encoding helper functions
+//===----------------------------------------------------------------------===//
+
+bool R600MCCodeEmitter::isFCOp(unsigned opcode) const {
+  switch(opcode) {
+  default: return false;
+  case AMDGPU::PREDICATED_BREAK:
+  case AMDGPU::CONTINUE:
+  case AMDGPU::IF_PREDICATE_SET:
+  case AMDGPU::ELSE:
+  case AMDGPU::ENDIF:
+  case AMDGPU::ENDLOOP:
+  case AMDGPU::WHILELOOP:
+    return true;
+  }
+}
+
+bool R600MCCodeEmitter::isTexOp(unsigned opcode) const {
+  switch(opcode) {
+  default: return false;
+  case AMDGPU::TEX_LD:
+  case AMDGPU::TEX_GET_TEXTURE_RESINFO:
+  case AMDGPU::TEX_SAMPLE:
+  case AMDGPU::TEX_SAMPLE_C:
+  case AMDGPU::TEX_SAMPLE_L:
+  case AMDGPU::TEX_SAMPLE_C_L:
+  case AMDGPU::TEX_SAMPLE_LB:
+  case AMDGPU::TEX_SAMPLE_C_LB:
+  case AMDGPU::TEX_SAMPLE_G:
+  case AMDGPU::TEX_SAMPLE_C_G:
+  case AMDGPU::TEX_GET_GRADIENTS_H:
+  case AMDGPU::TEX_GET_GRADIENTS_V:
+  case AMDGPU::TEX_SET_GRADIENTS_H:
+  case AMDGPU::TEX_SET_GRADIENTS_V:
+    return true;
+  }
+}
+
+bool R600MCCodeEmitter::isFlagSet(const MCInst &MI, unsigned Operand,
+                                  unsigned Flag) const {
+  const MCInstrDesc &MCDesc = MCII.get(MI.getOpcode());
+  unsigned FlagIndex = GET_FLAG_OPERAND_IDX(MCDesc.TSFlags);
+  if (FlagIndex == 0) {
+    return false;
+  }
+  assert(MI.getOperand(FlagIndex).isImm());
+  return !!((MI.getOperand(FlagIndex).getImm() >>
+            (NUM_MO_FLAGS * Operand)) & Flag);
+}
+
+#include "AMDGPUGenMCCodeEmitter.inc"
diff --git a/lib/Target/R600/MCTargetDesc/SIMCCodeEmitter.cpp b/lib/Target/R600/MCTargetDesc/SIMCCodeEmitter.cpp
new file mode 100644
index 0000000000..7f271d1c45
--- /dev/null
+++ b/lib/Target/R600/MCTargetDesc/SIMCCodeEmitter.cpp
@@ -0,0 +1,292 @@
+//===-- SIMCCodeEmitter.cpp - SI Code Emitter -------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief The SI code emitter produces machine code that can be executed
+/// directly on the GPU device.
+//
+//===----------------------------------------------------------------------===//
+
+#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
+#include "MCTargetDesc/AMDGPUMCCodeEmitter.h"
+#include "llvm/MC/MCCodeEmitter.h"
+#include "llvm/MC/MCContext.h"
+#include "llvm/MC/MCInst.h"
+#include "llvm/MC/MCInstrInfo.h"
+#include "llvm/MC/MCRegisterInfo.h"
+#include "llvm/MC/MCSubtargetInfo.h"
+#include "llvm/Support/raw_ostream.h"
+
+#define VGPR_BIT(src_idx) (1ULL << (9 * src_idx - 1))
+#define SI_INSTR_FLAGS_ENCODING_MASK 0xf
+
+// These must be kept in sync with SIInstructions.td and also the
+// InstrEncodingInfo array in SIInstrInfo.cpp.
+//
+// NOTE: This enum is only used to identify the encoding type within LLVM,
+// the actual encoding type that is part of the instruction format is different
+namespace SIInstrEncodingType {
+  enum Encoding {
+    EXP = 0,
+    LDS = 1,
+    MIMG = 2,
+    MTBUF = 3,
+    MUBUF = 4,
+    SMRD = 5,
+    SOP1 = 6,
+    SOP2 = 7,
+    SOPC = 8,
+    SOPK = 9,
+    SOPP = 10,
+    VINTRP = 11,
+    VOP1 = 12,
+    VOP2 = 13,
+    VOP3 = 14,
+    VOPC = 15
+  };
+}
+
+using namespace llvm;
+
+namespace {
+class SIMCCodeEmitter : public  AMDGPUMCCodeEmitter {
+  SIMCCodeEmitter(const SIMCCodeEmitter &); // DO NOT IMPLEMENT
+  void operator=(const SIMCCodeEmitter &); // DO NOT IMPLEMENT
+  const MCInstrInfo &MCII;
+  const MCRegisterInfo &MRI;
+  const MCSubtargetInfo &STI;
+  MCContext &Ctx;
+
+public:
+  SIMCCodeEmitter(const MCInstrInfo &mcii, const MCRegisterInfo &mri,
+                  const MCSubtargetInfo &sti, MCContext &ctx)
+    : MCII(mcii), MRI(mri), STI(sti), Ctx(ctx) { }
+
+  ~SIMCCodeEmitter() { }
+
+  /// \breif Encode the instruction and write it to the OS.
+  virtual void EncodeInstruction(const MCInst &MI, raw_ostream &OS,
+                         SmallVectorImpl<MCFixup> &Fixups) const;
+
+  /// \returns the encoding for an MCOperand.
+  virtual uint64_t getMachineOpValue(const MCInst &MI, const MCOperand &MO,
+                                     SmallVectorImpl<MCFixup> &Fixups) const;
+
+public:
+
+  /// \brief Encode a sequence of registers with the correct alignment.
+  unsigned GPRAlign(const MCInst &MI, unsigned OpNo, unsigned shift) const;
+
+  /// \brief Encoding for when 2 consecutive registers are used
+  virtual unsigned GPR2AlignEncode(const MCInst &MI, unsigned OpNo,
+                                   SmallVectorImpl<MCFixup> &Fixup) const;
+
+  /// \brief Encoding for when 4 consectuive registers are used
+  virtual unsigned GPR4AlignEncode(const MCInst &MI, unsigned OpNo,
+                                   SmallVectorImpl<MCFixup> &Fixup) const;
+
+  /// \brief Encoding for SMRD indexed loads
+  virtual uint32_t SMRDmemriEncode(const MCInst &MI, unsigned OpNo,
+                                   SmallVectorImpl<MCFixup> &Fixup) const;
+
+  /// \brief Post-Encoder method for VOP instructions
+  virtual uint64_t VOPPostEncode(const MCInst &MI, uint64_t Value) const;
+
+private:
+
+  /// \returns this SIInstrEncodingType for this instruction.
+  unsigned getEncodingType(const MCInst &MI) const;
+
+  /// \brief Get then size in bytes of this instructions encoding.
+  unsigned getEncodingBytes(const MCInst &MI) const;
+
+  /// \returns the hardware encoding for a register
+  unsigned getRegBinaryCode(unsigned reg) const;
+
+  /// \brief Generated function that returns the hardware encoding for
+  /// a register
+  unsigned getHWRegNum(unsigned reg) const;
+
+};
+
+} // End anonymous namespace
+
+MCCodeEmitter *llvm::createSIMCCodeEmitter(const MCInstrInfo &MCII,
+                                           const MCRegisterInfo &MRI,
+                                           const MCSubtargetInfo &STI,
+                                           MCContext &Ctx) {
+  return new SIMCCodeEmitter(MCII, MRI, STI, Ctx);
+}
+
+void SIMCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
+                                       SmallVectorImpl<MCFixup> &Fixups) const {
+  uint64_t Encoding = getBinaryCodeForInstr(MI, Fixups);
+  unsigned bytes = getEncodingBytes(MI);
+  for (unsigned i = 0; i < bytes; i++) {
+    OS.write((uint8_t) ((Encoding >> (8 * i)) & 0xff));
+  }
+}
+
+uint64_t SIMCCodeEmitter::getMachineOpValue(const MCInst &MI,
+                                            const MCOperand &MO,
+                                       SmallVectorImpl<MCFixup> &Fixups) const {
+  if (MO.isReg()) {
+    return getRegBinaryCode(MO.getReg());
+  } else if (MO.isImm()) {
+    return MO.getImm();
+  } else if (MO.isFPImm()) {
+    // XXX: Not all instructions can use inline literals
+    // XXX: We should make sure this is a 32-bit constant
+    union {
+      float F;
+      uint32_t I;
+    } Imm;
+    Imm.F = MO.getFPImm();
+    return Imm.I;
+  } else{
+    llvm_unreachable("Encoding of this operand type is not supported yet.");
+  }
+  return 0;
+}
+
+//===----------------------------------------------------------------------===//
+// Custom Operand Encodings
+//===----------------------------------------------------------------------===//
+
+unsigned SIMCCodeEmitter::GPRAlign(const MCInst &MI, unsigned OpNo,
+                                   unsigned shift) const {
+  unsigned regCode = getRegBinaryCode(MI.getOperand(OpNo).getReg());
+  return regCode >> shift;
+  return 0;
+}
+unsigned SIMCCodeEmitter::GPR2AlignEncode(const MCInst &MI,
+                                          unsigned OpNo ,
+                                        SmallVectorImpl<MCFixup> &Fixup) const {
+  return GPRAlign(MI, OpNo, 1);
+}
+
+unsigned SIMCCodeEmitter::GPR4AlignEncode(const MCInst &MI,
+                                          unsigned OpNo,
+                                        SmallVectorImpl<MCFixup> &Fixup) const {
+  return GPRAlign(MI, OpNo, 2);
+}
+
+#define SMRD_OFFSET_MASK 0xff
+#define SMRD_IMM_SHIFT 8
+#define SMRD_SBASE_MASK 0x3f
+#define SMRD_SBASE_SHIFT 9
+/// This function is responsibe for encoding the offset
+/// and the base ptr for SMRD instructions it should return a bit string in
+/// this format:
+///
+/// OFFSET = bits{7-0}
+/// IMM    = bits{8}
+/// SBASE  = bits{14-9}
+///
+uint32_t SIMCCodeEmitter::SMRDmemriEncode(const MCInst &MI, unsigned OpNo,
+                                        SmallVectorImpl<MCFixup> &Fixup) const {
+  uint32_t Encoding;
+
+  const MCOperand &OffsetOp = MI.getOperand(OpNo + 1);
+
+  //XXX: Use this function for SMRD loads with register offsets
+  assert(OffsetOp.isImm());
+
+  Encoding =
+      (getMachineOpValue(MI, OffsetOp, Fixup) & SMRD_OFFSET_MASK)
+    | (1 << SMRD_IMM_SHIFT) //XXX If the Offset is a register we shouldn't set this bit
+    | ((GPR2AlignEncode(MI, OpNo, Fixup) & SMRD_SBASE_MASK) << SMRD_SBASE_SHIFT)
+    ;
+
+  return Encoding;
+}
+
+//===----------------------------------------------------------------------===//
+// Post Encoder Callbacks
+//===----------------------------------------------------------------------===//
+
+uint64_t SIMCCodeEmitter::VOPPostEncode(const MCInst &MI, uint64_t Value) const{
+  unsigned encodingType = getEncodingType(MI);
+  unsigned numSrcOps;
+  unsigned vgprBitOffset;
+
+  if (encodingType == SIInstrEncodingType::VOP3) {
+    numSrcOps = 3;
+    vgprBitOffset = 32;
+  } else {
+    numSrcOps = 1;
+    vgprBitOffset = 0;
+  }
+
+  // Add one to skip over the destination reg operand.
+  for (unsigned opIdx = 1; opIdx < numSrcOps + 1; opIdx++) {
+    const MCOperand &MO = MI.getOperand(opIdx);
+    if (MO.isReg()) {
+      unsigned reg = MI.getOperand(opIdx).getReg();
+      if (AMDGPUMCRegisterClasses[AMDGPU::VReg_32RegClassID].contains(reg) ||
+          AMDGPUMCRegisterClasses[AMDGPU::VReg_64RegClassID].contains(reg)) {
+        Value |= (VGPR_BIT(opIdx)) << vgprBitOffset;
+      }
+    } else if (MO.isFPImm()) {
+      union {
+        float f;
+        uint32_t i;
+      } Imm;
+      // XXX: Not all instructions can use inline literals
+      // XXX: We should make sure this is a 32-bit constant
+      Imm.f = MO.getFPImm();
+      Value |= ((uint64_t)Imm.i) << 32;
+    }
+  }
+  return Value;
+}
+
+//===----------------------------------------------------------------------===//
+// Encoding helper functions
+//===----------------------------------------------------------------------===//
+
+unsigned SIMCCodeEmitter::getEncodingType(const MCInst &MI) const {
+  return MCII.get(MI.getOpcode()).TSFlags & SI_INSTR_FLAGS_ENCODING_MASK;
+}
+
+unsigned SIMCCodeEmitter::getEncodingBytes(const MCInst &MI) const {
+
+  // These instructions aren't real instructions with an encoding type, so
+  // we need to manually specify their size.
+  switch (MI.getOpcode()) {
+  default: break;
+  case AMDGPU::SI_LOAD_LITERAL_I32:
+  case AMDGPU::SI_LOAD_LITERAL_F32:
+    return 4;
+  }
+
+  unsigned encoding_type = getEncodingType(MI);
+  switch (encoding_type) {
+    case SIInstrEncodingType::EXP:
+    case SIInstrEncodingType::LDS:
+    case SIInstrEncodingType::MUBUF:
+    case SIInstrEncodingType::MTBUF:
+    case SIInstrEncodingType::MIMG:
+    case SIInstrEncodingType::VOP3:
+      return 8;
+    default:
+      return 4;
+  }
+}
+
+
+unsigned SIMCCodeEmitter::getRegBinaryCode(unsigned reg) const {
+  switch (reg) {
+    case AMDGPU::M0: return 124;
+    case AMDGPU::SREG_LIT_0: return 128;
+    case AMDGPU::SI_LITERAL_CONSTANT: return 255;
+    default: return MRI.getEncodingValue(reg);
+  }
+}
+
diff --git a/lib/Target/R600/Makefile b/lib/Target/R600/Makefile
new file mode 100644
index 0000000000..1b3ebbe8c8
--- /dev/null
+++ b/lib/Target/R600/Makefile
@@ -0,0 +1,23 @@
+##===- lib/Target/R600/Makefile ---------------------------*- Makefile -*-===##
+#
+#                     The LLVM Compiler Infrastructure
+#
+# This file is distributed under the University of Illinois Open Source
+# License. See LICENSE.TXT for details.
+#
+##===----------------------------------------------------------------------===##
+
+LEVEL = ../../..
+LIBRARYNAME = LLVMR600CodeGen
+TARGET = AMDGPU
+
+# Make sure that tblgen is run, first thing.
+BUILT_SOURCES = AMDGPUGenRegisterInfo.inc AMDGPUGenInstrInfo.inc \
+		AMDGPUGenDAGISel.inc  AMDGPUGenSubtargetInfo.inc \
+		AMDGPUGenMCCodeEmitter.inc AMDGPUGenCallingConv.inc \
+		AMDGPUGenIntrinsics.inc AMDGPUGenDFAPacketizer.inc \
+		AMDGPUGenAsmWriter.inc
+
+DIRS = InstPrinter TargetInfo MCTargetDesc
+
+include $(LEVEL)/Makefile.common
diff --git a/lib/Target/R600/Processors.td b/lib/Target/R600/Processors.td
new file mode 100644
index 0000000000..3dc1ecda77
--- /dev/null
+++ b/lib/Target/R600/Processors.td
@@ -0,0 +1,29 @@
+//===-- Processors.td - TODO: Add brief description -------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// AMDIL processors supported.
+//
+//===----------------------------------------------------------------------===//
+
+class Proc<string Name, ProcessorItineraries itin, list<SubtargetFeature> Features>
+: Processor<Name, itin, Features>;
+def : Proc<"r600",       R600_EG_Itin, [FeatureR600ALUInst]>;
+def : Proc<"rv710",      R600_EG_Itin, []>;
+def : Proc<"rv730",      R600_EG_Itin, []>;
+def : Proc<"rv770",      R600_EG_Itin, [FeatureFP64]>;
+def : Proc<"cedar",      R600_EG_Itin, [FeatureByteAddress, FeatureImages]>;
+def : Proc<"redwood",    R600_EG_Itin, [FeatureByteAddress, FeatureImages]>;
+def : Proc<"juniper",    R600_EG_Itin, [FeatureByteAddress, FeatureImages]>;
+def : Proc<"cypress",    R600_EG_Itin, [FeatureByteAddress, FeatureImages, FeatureFP64]>;
+def : Proc<"barts",      R600_EG_Itin, [FeatureByteAddress, FeatureImages]>;
+def : Proc<"turks",      R600_EG_Itin, [FeatureByteAddress, FeatureImages]>;
+def : Proc<"caicos",     R600_EG_Itin, [FeatureByteAddress, FeatureImages]>;
+def : Proc<"cayman",     R600_EG_Itin, [FeatureByteAddress, FeatureImages, FeatureFP64]>;
+def : Proc<"SI", SI_Itin, [Feature64BitPtr]>;
+
diff --git a/lib/Target/R600/R600Defines.h b/lib/Target/R600/R600Defines.h
new file mode 100644
index 0000000000..7dea8e44ea
--- /dev/null
+++ b/lib/Target/R600/R600Defines.h
@@ -0,0 +1,79 @@
+//===-- R600Defines.h - R600 Helper Macros ----------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//===----------------------------------------------------------------------===//
+
+#ifndef R600DEFINES_H_
+#define R600DEFINES_H_
+
+#include "llvm/MC/MCRegisterInfo.h"
+
+// Operand Flags
+#define MO_FLAG_CLAMP (1 << 0)
+#define MO_FLAG_NEG   (1 << 1)
+#define MO_FLAG_ABS   (1 << 2)
+#define MO_FLAG_MASK  (1 << 3)
+#define MO_FLAG_PUSH  (1 << 4)
+#define MO_FLAG_NOT_LAST  (1 << 5)
+#define MO_FLAG_LAST  (1 << 6)
+#define NUM_MO_FLAGS 7
+
+/// \brief Helper for getting the operand index for the instruction flags
+/// operand.
+#define GET_FLAG_OPERAND_IDX(Flags) (((Flags) >> 7) & 0x3)
+
+namespace R600_InstFlag {
+  enum TIF {
+    TRANS_ONLY = (1 << 0),
+    TEX = (1 << 1),
+    REDUCTION = (1 << 2),
+    FC = (1 << 3),
+    TRIG = (1 << 4),
+    OP3 = (1 << 5),
+    VECTOR = (1 << 6),
+    //FlagOperand bits 7, 8
+    NATIVE_OPERANDS = (1 << 9),
+    OP1 = (1 << 10),
+    OP2 = (1 << 11)
+  };
+}
+
+#define HAS_NATIVE_OPERANDS(Flags) ((Flags) & R600_InstFlag::NATIVE_OPERANDS)
+
+/// \brief Defines for extracting register infomation from register encoding
+#define HW_REG_MASK 0x1ff
+#define HW_CHAN_SHIFT 9
+
+namespace R600Operands {
+  enum Ops {
+    DST,
+    UPDATE_EXEC_MASK,
+    UPDATE_PREDICATE,
+    WRITE,
+    OMOD,
+    DST_REL,
+    CLAMP,
+    SRC0,
+    SRC0_NEG,
+    SRC0_REL,
+    SRC0_ABS,
+    SRC1,
+    SRC1_NEG,
+    SRC1_REL,
+    SRC1_ABS,
+    SRC2,
+    SRC2_NEG,
+    SRC2_REL,
+    LAST,
+    PRED_SEL,
+    IMM,
+    COUNT
+ };
+}
+
+#endif // R600DEFINES_H_
diff --git a/lib/Target/R600/R600ExpandSpecialInstrs.cpp b/lib/Target/R600/R600ExpandSpecialInstrs.cpp
new file mode 100644
index 0000000000..58221f9ec2
--- /dev/null
+++ b/lib/Target/R600/R600ExpandSpecialInstrs.cpp
@@ -0,0 +1,333 @@
+//===-- R600ExpandSpecialInstrs.cpp - Expand special instructions ---------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// Vector, Reduction, and Cube instructions need to fill the entire instruction
+/// group to work correctly.  This pass expands these individual instructions
+/// into several instructions that will completely fill the instruction group.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "R600Defines.h"
+#include "R600InstrInfo.h"
+#include "R600RegisterInfo.h"
+#include "R600MachineFunctionInfo.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+
+using namespace llvm;
+
+namespace {
+
+class R600ExpandSpecialInstrsPass : public MachineFunctionPass {
+
+private:
+  static char ID;
+  const R600InstrInfo *TII;
+
+  bool ExpandInputPerspective(MachineInstr& MI);
+  bool ExpandInputConstant(MachineInstr& MI);
+
+public:
+  R600ExpandSpecialInstrsPass(TargetMachine &tm) : MachineFunctionPass(ID),
+    TII (static_cast<const R600InstrInfo *>(tm.getInstrInfo())) { }
+
+  virtual bool runOnMachineFunction(MachineFunction &MF);
+
+  const char *getPassName() const {
+    return "R600 Expand special instructions pass";
+  }
+};
+
+} // End anonymous namespace
+
+char R600ExpandSpecialInstrsPass::ID = 0;
+
+FunctionPass *llvm::createR600ExpandSpecialInstrsPass(TargetMachine &TM) {
+  return new R600ExpandSpecialInstrsPass(TM);
+}
+
+bool R600ExpandSpecialInstrsPass::ExpandInputPerspective(MachineInstr &MI) {
+  const R600RegisterInfo &TRI = TII->getRegisterInfo();
+  if (MI.getOpcode() != AMDGPU::input_perspective)
+    return false;
+
+  MachineBasicBlock::iterator I = &MI;
+  unsigned DstReg = MI.getOperand(0).getReg();
+  R600MachineFunctionInfo *MFI = MI.getParent()->getParent()
+      ->getInfo<R600MachineFunctionInfo>();
+  unsigned IJIndexBase;
+
+  // In Evergreen ISA doc section 8.3.2 :
+  // We need to interpolate XY and ZW in two different instruction groups.
+  // An INTERP_* must occupy all 4 slots of an instruction group.
+  // Output of INTERP_XY is written in X,Y slots
+  // Output of INTERP_ZW is written in Z,W slots
+  //
+  // Thus interpolation requires the following sequences :
+  //
+  // AnyGPR.x = INTERP_ZW; (Write Masked Out)
+  // AnyGPR.y = INTERP_ZW; (Write Masked Out)
+  // DstGPR.z = INTERP_ZW;
+  // DstGPR.w = INTERP_ZW; (End of first IG)
+  // DstGPR.x = INTERP_XY;
+  // DstGPR.y = INTERP_XY;
+  // AnyGPR.z = INTERP_XY; (Write Masked Out)
+  // AnyGPR.w = INTERP_XY; (Write Masked Out) (End of second IG)
+  //
+  switch (MI.getOperand(1).getImm()) {
+  case 0:
+    IJIndexBase = MFI->GetIJPerspectiveIndex();
+    break;
+  case 1:
+    IJIndexBase = MFI->GetIJLinearIndex();
+    break;
+  default:
+    assert(0 && "Unknow ij index");
+  }
+
+  for (unsigned i = 0; i < 8; i++) {
+    unsigned IJIndex = AMDGPU::R600_TReg32RegClass.getRegister(
+        2 * IJIndexBase + ((i + 1) % 2));
+    unsigned ReadReg = AMDGPU::R600_ArrayBaseRegClass.getRegister(
+        MI.getOperand(2).getImm());
+
+
+    unsigned Sel = AMDGPU::sel_x;
+    switch (i % 4) {
+    case 0:Sel = AMDGPU::sel_x;break;
+    case 1:Sel = AMDGPU::sel_y;break;
+    case 2:Sel = AMDGPU::sel_z;break;
+    case 3:Sel = AMDGPU::sel_w;break;
+    default:break;
+    }
+
+    unsigned Res = TRI.getSubReg(DstReg, Sel);
+
+    unsigned Opcode = (i < 4)?AMDGPU::INTERP_ZW:AMDGPU::INTERP_XY;
+
+    MachineBasicBlock &MBB = *(MI.getParent());
+    MachineInstr *NewMI =
+        TII->buildDefaultInstruction(MBB, I, Opcode, Res, IJIndex, ReadReg);
+
+    if (!(i> 1 && i < 6)) {
+      TII->addFlag(NewMI, 0, MO_FLAG_MASK);
+    }
+
+    if (i % 4 !=  3)
+      TII->addFlag(NewMI, 0, MO_FLAG_NOT_LAST);
+  }
+
+  MI.eraseFromParent();
+
+  return true;
+}
+
+bool R600ExpandSpecialInstrsPass::ExpandInputConstant(MachineInstr &MI) {
+  const R600RegisterInfo &TRI = TII->getRegisterInfo();
+  if (MI.getOpcode() != AMDGPU::input_constant)
+    return false;
+
+  MachineBasicBlock::iterator I = &MI;
+  unsigned DstReg = MI.getOperand(0).getReg();
+
+  for (unsigned i = 0; i < 4; i++) {
+    unsigned ReadReg = AMDGPU::R600_ArrayBaseRegClass.getRegister(
+        MI.getOperand(1).getImm());
+
+    unsigned Sel = AMDGPU::sel_x;
+    switch (i % 4) {
+    case 0:Sel = AMDGPU::sel_x;break;
+    case 1:Sel = AMDGPU::sel_y;break;
+    case 2:Sel = AMDGPU::sel_z;break;
+    case 3:Sel = AMDGPU::sel_w;break;
+    default:break;
+    }
+
+    unsigned Res = TRI.getSubReg(DstReg, Sel);
+
+    MachineBasicBlock &MBB = *(MI.getParent());
+    MachineInstr *NewMI = TII->buildDefaultInstruction(
+        MBB, I, AMDGPU::INTERP_LOAD_P0, Res, ReadReg);
+
+    if (i % 4 !=  3)
+      TII->addFlag(NewMI, 0, MO_FLAG_NOT_LAST);
+  }
+
+  MI.eraseFromParent();
+
+  return true;
+}
+
+bool R600ExpandSpecialInstrsPass::runOnMachineFunction(MachineFunction &MF) {
+
+  const R600RegisterInfo &TRI = TII->getRegisterInfo();
+
+  for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
+                                                  BB != BB_E; ++BB) {
+    MachineBasicBlock &MBB = *BB;
+    MachineBasicBlock::iterator I = MBB.begin();
+    while (I != MBB.end()) {
+      MachineInstr &MI = *I;
+      I = llvm::next(I);
+
+      switch (MI.getOpcode()) {
+      default: break;
+      // Expand PRED_X to one of the PRED_SET instructions.
+      case AMDGPU::PRED_X: {
+        uint64_t Flags = MI.getOperand(3).getImm();
+        // The native opcode used by PRED_X is stored as an immediate in the
+        // third operand.
+        MachineInstr *PredSet = TII->buildDefaultInstruction(MBB, I,
+                                            MI.getOperand(2).getImm(), // opcode
+                                            MI.getOperand(0).getReg(), // dst
+                                            MI.getOperand(1).getReg(), // src0
+                                            AMDGPU::ZERO);             // src1
+        TII->addFlag(PredSet, 0, MO_FLAG_MASK);
+        if (Flags & MO_FLAG_PUSH) {
+          TII->setImmOperand(PredSet, R600Operands::UPDATE_EXEC_MASK, 1);
+        } else {
+          TII->setImmOperand(PredSet, R600Operands::UPDATE_PREDICATE, 1);
+        }
+        MI.eraseFromParent();
+        continue;
+        }
+      case AMDGPU::BREAK:
+        MachineInstr *PredSet = TII->buildDefaultInstruction(MBB, I,
+                                          AMDGPU::PRED_SETE_INT,
+                                          AMDGPU::PREDICATE_BIT,
+                                          AMDGPU::ZERO,
+                                          AMDGPU::ZERO);
+        TII->addFlag(PredSet, 0, MO_FLAG_MASK);
+        TII->setImmOperand(PredSet, R600Operands::UPDATE_EXEC_MASK, 1);
+
+        BuildMI(MBB, I, MBB.findDebugLoc(I),
+                TII->get(AMDGPU::PREDICATED_BREAK))
+                .addReg(AMDGPU::PREDICATE_BIT);
+        MI.eraseFromParent();
+        continue;
+    }
+
+    if (ExpandInputPerspective(MI))
+      continue;
+    if (ExpandInputConstant(MI))
+      continue;
+
+      bool IsReduction = TII->isReductionOp(MI.getOpcode());
+      bool IsVector = TII->isVector(MI);
+      bool IsCube = TII->isCubeOp(MI.getOpcode());
+      if (!IsReduction && !IsVector && !IsCube) {
+        continue;
+      }
+
+      // Expand the instruction
+      //
+      // Reduction instructions:
+      // T0_X = DP4 T1_XYZW, T2_XYZW
+      // becomes:
+      // TO_X = DP4 T1_X, T2_X
+      // TO_Y (write masked) = DP4 T1_Y, T2_Y
+      // TO_Z (write masked) = DP4 T1_Z, T2_Z
+      // TO_W (write masked) = DP4 T1_W, T2_W
+      //
+      // Vector instructions:
+      // T0_X = MULLO_INT T1_X, T2_X
+      // becomes:
+      // T0_X = MULLO_INT T1_X, T2_X
+      // T0_Y (write masked) = MULLO_INT T1_X, T2_X
+      // T0_Z (write masked) = MULLO_INT T1_X, T2_X
+      // T0_W (write masked) = MULLO_INT T1_X, T2_X
+      //
+      // Cube instructions:
+      // T0_XYZW = CUBE T1_XYZW
+      // becomes:
+      // TO_X = CUBE T1_Z, T1_Y
+      // T0_Y = CUBE T1_Z, T1_X
+      // T0_Z = CUBE T1_X, T1_Z
+      // T0_W = CUBE T1_Y, T1_Z
+      for (unsigned Chan = 0; Chan < 4; Chan++) {
+        unsigned DstReg = MI.getOperand(
+                            TII->getOperandIdx(MI, R600Operands::DST)).getReg();
+        unsigned Src0 = MI.getOperand(
+                           TII->getOperandIdx(MI, R600Operands::SRC0)).getReg();
+        unsigned Src1 = 0;
+
+        // Determine the correct source registers
+        if (!IsCube) {
+          int Src1Idx = TII->getOperandIdx(MI, R600Operands::SRC1);
+          if (Src1Idx != -1) {
+            Src1 = MI.getOperand(Src1Idx).getReg();
+          }
+        }
+        if (IsReduction) {
+          unsigned SubRegIndex = TRI.getSubRegFromChannel(Chan);
+          Src0 = TRI.getSubReg(Src0, SubRegIndex);
+          Src1 = TRI.getSubReg(Src1, SubRegIndex);
+        } else if (IsCube) {
+          static const int CubeSrcSwz[] = {2, 2, 0, 1};
+          unsigned SubRegIndex0 = TRI.getSubRegFromChannel(CubeSrcSwz[Chan]);
+          unsigned SubRegIndex1 = TRI.getSubRegFromChannel(CubeSrcSwz[3 - Chan]);
+          Src1 = TRI.getSubReg(Src0, SubRegIndex1);
+          Src0 = TRI.getSubReg(Src0, SubRegIndex0);
+        }
+
+        // Determine the correct destination registers;
+        bool Mask = false;
+        bool NotLast = true;
+        if (IsCube) {
+          unsigned SubRegIndex = TRI.getSubRegFromChannel(Chan);
+          DstReg = TRI.getSubReg(DstReg, SubRegIndex);
+        } else {
+          // Mask the write if the original instruction does not write to
+          // the current Channel.
+          Mask = (Chan != TRI.getHWRegChan(DstReg));
+          unsigned DstBase = TRI.getEncodingValue(DstReg) & HW_REG_MASK;
+          DstReg = AMDGPU::R600_TReg32RegClass.getRegister((DstBase * 4) + Chan);
+        }
+
+        // Set the IsLast bit
+        NotLast = (Chan != 3 );
+
+        // Add the new instruction
+        unsigned Opcode = MI.getOpcode();
+        switch (Opcode) {
+        case AMDGPU::CUBE_r600_pseudo:
+          Opcode = AMDGPU::CUBE_r600_real;
+          break;
+        case AMDGPU::CUBE_eg_pseudo:
+          Opcode = AMDGPU::CUBE_eg_real;
+          break;
+        case AMDGPU::DOT4_r600_pseudo:
+          Opcode = AMDGPU::DOT4_r600_real;
+          break;
+        case AMDGPU::DOT4_eg_pseudo:
+          Opcode = AMDGPU::DOT4_eg_real;
+          break;
+        default:
+          break;
+        }
+
+        MachineInstr *NewMI =
+          TII->buildDefaultInstruction(MBB, I, Opcode, DstReg, Src0, Src1);
+
+        NewMI->setIsInsideBundle(Chan != 0);
+        if (Mask) {
+          TII->addFlag(NewMI, 0, MO_FLAG_MASK);
+        }
+        if (NotLast) {
+          TII->addFlag(NewMI, 0, MO_FLAG_NOT_LAST);
+        }
+      }
+      MI.eraseFromParent();
+    }
+  }
+  return false;
+}
diff --git a/lib/Target/R600/R600ISelLowering.cpp b/lib/Target/R600/R600ISelLowering.cpp
new file mode 100644
index 0000000000..eaeff4ec21
--- /dev/null
+++ b/lib/Target/R600/R600ISelLowering.cpp
@@ -0,0 +1,905 @@
+//===-- R600ISelLowering.cpp - R600 DAG Lowering Implementation -----------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Custom DAG lowering for R600
+//
+//===----------------------------------------------------------------------===//
+
+#include "R600ISelLowering.h"
+#include "R600Defines.h"
+#include "R600InstrInfo.h"
+#include "R600MachineFunctionInfo.h"
+#include "llvm/Argument.h"
+#include "llvm/Function.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/SelectionDAG.h"
+
+using namespace llvm;
+
+R600TargetLowering::R600TargetLowering(TargetMachine &TM) :
+    AMDGPUTargetLowering(TM),
+    TII(static_cast<const R600InstrInfo*>(TM.getInstrInfo())) {
+  setOperationAction(ISD::MUL, MVT::i64, Expand);
+  addRegisterClass(MVT::v4f32, &AMDGPU::R600_Reg128RegClass);
+  addRegisterClass(MVT::f32, &AMDGPU::R600_Reg32RegClass);
+  addRegisterClass(MVT::v4i32, &AMDGPU::R600_Reg128RegClass);
+  addRegisterClass(MVT::i32, &AMDGPU::R600_Reg32RegClass);
+  computeRegisterProperties();
+
+  setOperationAction(ISD::FADD, MVT::v4f32, Expand);
+  setOperationAction(ISD::FMUL, MVT::v4f32, Expand);
+  setOperationAction(ISD::FDIV, MVT::v4f32, Expand);
+  setOperationAction(ISD::FSUB, MVT::v4f32, Expand);
+
+  setOperationAction(ISD::ADD,  MVT::v4i32, Expand);
+  setOperationAction(ISD::AND,  MVT::v4i32, Expand);
+  setOperationAction(ISD::UDIV, MVT::v4i32, Expand);
+  setOperationAction(ISD::UREM, MVT::v4i32, Expand);
+  setOperationAction(ISD::SETCC, MVT::v4i32, Expand);
+
+  setOperationAction(ISD::BR_CC, MVT::i32, Custom);
+  setOperationAction(ISD::BR_CC, MVT::f32, Custom);
+
+  setOperationAction(ISD::FSUB, MVT::f32, Expand);
+
+  setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
+  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
+  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i1, Custom);
+  setOperationAction(ISD::FPOW, MVT::f32, Custom);
+
+  setOperationAction(ISD::ROTL, MVT::i32, Custom);
+
+  setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
+  setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
+
+  setOperationAction(ISD::SETCC, MVT::i32, Custom);
+  setOperationAction(ISD::SETCC, MVT::f32, Custom);
+  setOperationAction(ISD::FP_TO_UINT, MVT::i1, Custom);
+
+  setOperationAction(ISD::SELECT, MVT::i32, Custom);
+  setOperationAction(ISD::SELECT, MVT::f32, Custom);
+
+  setOperationAction(ISD::STORE, MVT::i32, Custom);
+  setOperationAction(ISD::STORE, MVT::v4i32, Custom);
+
+  setTargetDAGCombine(ISD::FP_ROUND);
+
+  setSchedulingPreference(Sched::VLIW);
+}
+
+MachineBasicBlock * R600TargetLowering::EmitInstrWithCustomInserter(
+    MachineInstr * MI, MachineBasicBlock * BB) const {
+  MachineFunction * MF = BB->getParent();
+  MachineRegisterInfo &MRI = MF->getRegInfo();
+  MachineBasicBlock::iterator I = *MI;
+
+  switch (MI->getOpcode()) {
+  default: return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
+  case AMDGPU::SHADER_TYPE: break;
+  case AMDGPU::CLAMP_R600: {
+    MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
+                                                   AMDGPU::MOV,
+                                                   MI->getOperand(0).getReg(),
+                                                   MI->getOperand(1).getReg());
+    TII->addFlag(NewMI, 0, MO_FLAG_CLAMP);
+    break;
+  }
+
+  case AMDGPU::FABS_R600: {
+    MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
+                                                    AMDGPU::MOV,
+                                                    MI->getOperand(0).getReg(),
+                                                    MI->getOperand(1).getReg());
+    TII->addFlag(NewMI, 0, MO_FLAG_ABS);
+    break;
+  }
+
+  case AMDGPU::FNEG_R600: {
+    MachineInstr *NewMI = TII->buildDefaultInstruction(*BB, I,
+                                                    AMDGPU::MOV,
+                                                    MI->getOperand(0).getReg(),
+                                                    MI->getOperand(1).getReg());
+    TII->addFlag(NewMI, 0, MO_FLAG_NEG);
+    break;
+  }
+
+  case AMDGPU::R600_LOAD_CONST: {
+    int64_t RegIndex = MI->getOperand(1).getImm();
+    unsigned ConstantReg = AMDGPU::R600_CReg32RegClass.getRegister(RegIndex);
+    BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::COPY))
+                .addOperand(MI->getOperand(0))
+                .addReg(ConstantReg);
+    break;
+  }
+
+  case AMDGPU::MASK_WRITE: {
+    unsigned maskedRegister = MI->getOperand(0).getReg();
+    assert(TargetRegisterInfo::isVirtualRegister(maskedRegister));
+    MachineInstr * defInstr = MRI.getVRegDef(maskedRegister);
+    TII->addFlag(defInstr, 0, MO_FLAG_MASK);
+    break;
+  }
+
+  case AMDGPU::MOV_IMM_F32:
+    TII->buildMovImm(*BB, I, MI->getOperand(0).getReg(),
+                     MI->getOperand(1).getFPImm()->getValueAPF()
+                         .bitcastToAPInt().getZExtValue());
+    break;
+  case AMDGPU::MOV_IMM_I32:
+    TII->buildMovImm(*BB, I, MI->getOperand(0).getReg(),
+                     MI->getOperand(1).getImm());
+    break;
+
+
+  case AMDGPU::RAT_WRITE_CACHELESS_32_eg:
+  case AMDGPU::RAT_WRITE_CACHELESS_128_eg: {
+    unsigned EOP = (llvm::next(I)->getOpcode() == AMDGPU::RETURN) ? 1 : 0;
+
+    BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI->getOpcode()))
+            .addOperand(MI->getOperand(0))
+            .addOperand(MI->getOperand(1))
+            .addImm(EOP); // Set End of program bit
+    break;
+  }
+
+  case AMDGPU::RESERVE_REG: {
+    R600MachineFunctionInfo * MFI = MF->getInfo<R600MachineFunctionInfo>();
+    int64_t ReservedIndex = MI->getOperand(0).getImm();
+    unsigned ReservedReg =
+                         AMDGPU::R600_TReg32RegClass.getRegister(ReservedIndex);
+    MFI->ReservedRegs.push_back(ReservedReg);
+    unsigned SuperReg =
+          AMDGPU::R600_Reg128RegClass.getRegister(ReservedIndex / 4);
+    MFI->ReservedRegs.push_back(SuperReg);
+    break;
+  }
+
+  case AMDGPU::TXD: {
+    unsigned T0 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
+    unsigned T1 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
+
+    BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_H), T0)
+            .addOperand(MI->getOperand(3))
+            .addOperand(MI->getOperand(4))
+            .addOperand(MI->getOperand(5))
+            .addOperand(MI->getOperand(6));
+    BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_V), T1)
+            .addOperand(MI->getOperand(2))
+            .addOperand(MI->getOperand(4))
+            .addOperand(MI->getOperand(5))
+            .addOperand(MI->getOperand(6));
+    BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SAMPLE_G))
+            .addOperand(MI->getOperand(0))
+            .addOperand(MI->getOperand(1))
+            .addOperand(MI->getOperand(4))
+            .addOperand(MI->getOperand(5))
+            .addOperand(MI->getOperand(6))
+            .addReg(T0, RegState::Implicit)
+            .addReg(T1, RegState::Implicit);
+    break;
+  }
+
+  case AMDGPU::TXD_SHADOW: {
+    unsigned T0 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
+    unsigned T1 = MRI.createVirtualRegister(&AMDGPU::R600_Reg128RegClass);
+
+    BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_H), T0)
+            .addOperand(MI->getOperand(3))
+            .addOperand(MI->getOperand(4))
+            .addOperand(MI->getOperand(5))
+            .addOperand(MI->getOperand(6));
+    BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SET_GRADIENTS_V), T1)
+            .addOperand(MI->getOperand(2))
+            .addOperand(MI->getOperand(4))
+            .addOperand(MI->getOperand(5))
+            .addOperand(MI->getOperand(6));
+    BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::TEX_SAMPLE_C_G))
+            .addOperand(MI->getOperand(0))
+            .addOperand(MI->getOperand(1))
+            .addOperand(MI->getOperand(4))
+            .addOperand(MI->getOperand(5))
+            .addOperand(MI->getOperand(6))
+            .addReg(T0, RegState::Implicit)
+            .addReg(T1, RegState::Implicit);
+    break;
+  }
+
+  case AMDGPU::BRANCH:
+      BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP))
+              .addOperand(MI->getOperand(0))
+              .addReg(0);
+      break;
+
+  case AMDGPU::BRANCH_COND_f32: {
+    MachineInstr *NewMI =
+      BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::PRED_X),
+              AMDGPU::PREDICATE_BIT)
+              .addOperand(MI->getOperand(1))
+              .addImm(OPCODE_IS_NOT_ZERO)
+              .addImm(0); // Flags
+    TII->addFlag(NewMI, 0, MO_FLAG_PUSH);
+    BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP))
+            .addOperand(MI->getOperand(0))
+            .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
+    break;
+  }
+
+  case AMDGPU::BRANCH_COND_i32: {
+    MachineInstr *NewMI =
+      BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::PRED_X),
+            AMDGPU::PREDICATE_BIT)
+            .addOperand(MI->getOperand(1))
+            .addImm(OPCODE_IS_NOT_ZERO_INT)
+            .addImm(0); // Flags
+    TII->addFlag(NewMI, 0, MO_FLAG_PUSH);
+    BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::JUMP))
+           .addOperand(MI->getOperand(0))
+            .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
+    break;
+  }
+
+  case AMDGPU::input_perspective: {
+    R600MachineFunctionInfo *MFI = MF->getInfo<R600MachineFunctionInfo>();
+
+    // XXX Be more fine about register reservation
+    for (unsigned i = 0; i < 4; i ++) {
+      unsigned ReservedReg = AMDGPU::R600_TReg32RegClass.getRegister(i);
+      MFI->ReservedRegs.push_back(ReservedReg);
+    }
+
+    switch (MI->getOperand(1).getImm()) {
+    case 0:// Perspective
+      MFI->HasPerspectiveInterpolation = true;
+      break;
+    case 1:// Linear
+      MFI->HasLinearInterpolation = true;
+      break;
+    default:
+      assert(0 && "Unknow ij index");
+    }
+
+    return BB;
+  }
+
+  case AMDGPU::EG_ExportSwz:
+  case AMDGPU::R600_ExportSwz: {
+    bool EOP = (llvm::next(I)->getOpcode() == AMDGPU::RETURN)? 1 : 0;
+    if (!EOP)
+      return BB;
+    unsigned CfInst = (MI->getOpcode() == AMDGPU::EG_ExportSwz)? 84 : 40;
+    BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(MI->getOpcode()))
+            .addOperand(MI->getOperand(0))
+            .addOperand(MI->getOperand(1))
+            .addOperand(MI->getOperand(2))
+            .addOperand(MI->getOperand(3))
+            .addOperand(MI->getOperand(4))
+            .addOperand(MI->getOperand(5))
+            .addOperand(MI->getOperand(6))
+            .addImm(CfInst)
+            .addImm(1);
+    break;
+  }
+  }
+
+  MI->eraseFromParent();
+  return BB;
+}
+
+//===----------------------------------------------------------------------===//
+// Custom DAG Lowering Operations
+//===----------------------------------------------------------------------===//
+
+using namespace llvm::Intrinsic;
+using namespace llvm::AMDGPUIntrinsic;
+
+static SDValue
+InsertScalarToRegisterExport(SelectionDAG &DAG, DebugLoc DL, SDNode **ExportMap,
+    unsigned Slot, unsigned Channel, unsigned Inst, unsigned Type,
+    SDValue Scalar, SDValue Chain) {
+  if (!ExportMap[Slot]) {
+    SDValue Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT,
+      DL, MVT::v4f32,
+      DAG.getUNDEF(MVT::v4f32),
+      Scalar,
+      DAG.getConstant(Channel, MVT::i32));
+
+    unsigned Mask = 1 << Channel;
+
+    const SDValue Ops[] = {Chain, Vector, DAG.getConstant(Inst, MVT::i32),
+        DAG.getConstant(Type, MVT::i32), DAG.getConstant(Slot, MVT::i32),
+        DAG.getConstant(Mask, MVT::i32)};
+
+    SDValue Res =  DAG.getNode(
+        AMDGPUISD::EXPORT,
+        DL,
+        MVT::Other,
+        Ops, 6);
+     ExportMap[Slot] = Res.getNode();
+     return Res;
+  }
+
+  SDNode *ExportInstruction = (SDNode *) ExportMap[Slot] ;
+  SDValue PreviousVector = ExportInstruction->getOperand(1);
+  SDValue Vector = DAG.getNode(ISD::INSERT_VECTOR_ELT,
+      DL, MVT::v4f32,
+      PreviousVector,
+      Scalar,
+      DAG.getConstant(Channel, MVT::i32));
+
+  unsigned Mask = dyn_cast<ConstantSDNode>(ExportInstruction->getOperand(5))
+      ->getZExtValue();
+  Mask |= (1 << Channel);
+
+  const SDValue Ops[] = {ExportInstruction->getOperand(0), Vector,
+      DAG.getConstant(Inst, MVT::i32),
+      DAG.getConstant(Type, MVT::i32),
+      DAG.getConstant(Slot, MVT::i32),
+      DAG.getConstant(Mask, MVT::i32)};
+
+  DAG.UpdateNodeOperands(ExportInstruction,
+      Ops, 6);
+
+  return Chain;
+
+}
+
+SDValue R600TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
+  switch (Op.getOpcode()) {
+  default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
+  case ISD::BR_CC: return LowerBR_CC(Op, DAG);
+  case ISD::ROTL: return LowerROTL(Op, DAG);
+  case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
+  case ISD::SELECT: return LowerSELECT(Op, DAG);
+  case ISD::SETCC: return LowerSETCC(Op, DAG);
+  case ISD::STORE: return LowerSTORE(Op, DAG);
+  case ISD::FPOW: return LowerFPOW(Op, DAG);
+  case ISD::INTRINSIC_VOID: {
+    SDValue Chain = Op.getOperand(0);
+    unsigned IntrinsicID =
+                         cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
+    switch (IntrinsicID) {
+    case AMDGPUIntrinsic::AMDGPU_store_output: {
+      MachineFunction &MF = DAG.getMachineFunction();
+      MachineRegisterInfo &MRI = MF.getRegInfo();
+      int64_t RegIndex = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
+      unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister(RegIndex);
+      if (!MRI.isLiveOut(Reg)) {
+        MRI.addLiveOut(Reg);
+      }
+      return DAG.getCopyToReg(Chain, Op.getDebugLoc(), Reg, Op.getOperand(2));
+    }
+    case AMDGPUIntrinsic::R600_store_pixel_color: {
+      MachineFunction &MF = DAG.getMachineFunction();
+      R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
+      int64_t RegIndex = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
+
+      SDNode **OutputsMap = MFI->Outputs;
+      return InsertScalarToRegisterExport(DAG, Op.getDebugLoc(), OutputsMap,
+          RegIndex / 4, RegIndex % 4, 0, 0, Op.getOperand(2),
+          Chain);
+
+    }
+    case AMDGPUIntrinsic::R600_store_stream_output : {
+      MachineFunction &MF = DAG.getMachineFunction();
+      R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
+      int64_t RegIndex = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
+      int64_t BufIndex = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
+
+      SDNode **OutputsMap = MFI->StreamOutputs[BufIndex];
+      unsigned Inst;
+      switch (cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue()  ) {
+      // STREAM3
+      case 3:
+        Inst = 4;
+        break;
+      // STREAM2
+      case 2:
+        Inst = 3;
+        break;
+      // STREAM1
+      case 1:
+        Inst = 2;
+        break;
+      // STREAM0
+      case 0:
+        Inst = 1;
+        break;
+      default:
+        assert(0 && "Wrong buffer id for stream outputs !");
+      }
+
+      return InsertScalarToRegisterExport(DAG, Op.getDebugLoc(), OutputsMap,
+          RegIndex / 4, RegIndex % 4, Inst, 0, Op.getOperand(2),
+          Chain);
+    }
+    // default for switch(IntrinsicID)
+    default: break;
+    }
+    // break out of case ISD::INTRINSIC_VOID in switch(Op.getOpcode())
+    break;
+  }
+  case ISD::INTRINSIC_WO_CHAIN: {
+    unsigned IntrinsicID =
+                         cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
+    EVT VT = Op.getValueType();
+    DebugLoc DL = Op.getDebugLoc();
+    switch(IntrinsicID) {
+    default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
+    case AMDGPUIntrinsic::R600_load_input: {
+      int64_t RegIndex = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
+      unsigned Reg = AMDGPU::R600_TReg32RegClass.getRegister(RegIndex);
+      return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass, Reg, VT);
+    }
+    case AMDGPUIntrinsic::R600_load_input_perspective: {
+      int slot = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
+      if (slot < 0)
+        return DAG.getUNDEF(MVT::f32);
+      SDValue FullVector = DAG.getNode(
+          AMDGPUISD::INTERP,
+          DL, MVT::v4f32,
+          DAG.getConstant(0, MVT::i32), DAG.getConstant(slot / 4 , MVT::i32));
+      return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
+        DL, VT, FullVector, DAG.getConstant(slot % 4, MVT::i32));
+    }
+    case AMDGPUIntrinsic::R600_load_input_linear: {
+      int slot = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
+      if (slot < 0)
+        return DAG.getUNDEF(MVT::f32);
+      SDValue FullVector = DAG.getNode(
+        AMDGPUISD::INTERP,
+        DL, MVT::v4f32,
+        DAG.getConstant(1, MVT::i32), DAG.getConstant(slot / 4 , MVT::i32));
+      return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
+        DL, VT, FullVector, DAG.getConstant(slot % 4, MVT::i32));
+    }
+    case AMDGPUIntrinsic::R600_load_input_constant: {
+      int slot = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
+      if (slot < 0)
+        return DAG.getUNDEF(MVT::f32);
+      SDValue FullVector = DAG.getNode(
+        AMDGPUISD::INTERP_P0,
+        DL, MVT::v4f32,
+        DAG.getConstant(slot / 4 , MVT::i32));
+      return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
+          DL, VT, FullVector, DAG.getConstant(slot % 4, MVT::i32));
+    }
+
+    case r600_read_ngroups_x:
+      return LowerImplicitParameter(DAG, VT, DL, 0);
+    case r600_read_ngroups_y:
+      return LowerImplicitParameter(DAG, VT, DL, 1);
+    case r600_read_ngroups_z:
+      return LowerImplicitParameter(DAG, VT, DL, 2);
+    case r600_read_global_size_x:
+      return LowerImplicitParameter(DAG, VT, DL, 3);
+    case r600_read_global_size_y:
+      return LowerImplicitParameter(DAG, VT, DL, 4);
+    case r600_read_global_size_z:
+      return LowerImplicitParameter(DAG, VT, DL, 5);
+    case r600_read_local_size_x:
+      return LowerImplicitParameter(DAG, VT, DL, 6);
+    case r600_read_local_size_y:
+      return LowerImplicitParameter(DAG, VT, DL, 7);
+    case r600_read_local_size_z:
+      return LowerImplicitParameter(DAG, VT, DL, 8);
+
+    case r600_read_tgid_x:
+      return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
+                                  AMDGPU::T1_X, VT);
+    case r600_read_tgid_y:
+      return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
+                                  AMDGPU::T1_Y, VT);
+    case r600_read_tgid_z:
+      return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
+                                  AMDGPU::T1_Z, VT);
+    case r600_read_tidig_x:
+      return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
+                                  AMDGPU::T0_X, VT);
+    case r600_read_tidig_y:
+      return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
+                                  AMDGPU::T0_Y, VT);
+    case r600_read_tidig_z:
+      return CreateLiveInRegister(DAG, &AMDGPU::R600_TReg32RegClass,
+                                  AMDGPU::T0_Z, VT);
+    }
+    // break out of case ISD::INTRINSIC_WO_CHAIN in switch(Op.getOpcode())
+    break;
+  }
+  } // end switch(Op.getOpcode())
+  return SDValue();
+}
+
+void R600TargetLowering::ReplaceNodeResults(SDNode *N,
+                                            SmallVectorImpl<SDValue> &Results,
+                                            SelectionDAG &DAG) const {
+  switch (N->getOpcode()) {
+  default: return;
+  case ISD::FP_TO_UINT: Results.push_back(LowerFPTOUINT(N->getOperand(0), DAG));
+  }
+}
+
+SDValue R600TargetLowering::LowerFPTOUINT(SDValue Op, SelectionDAG &DAG) const {
+  return DAG.getNode(
+      ISD::SETCC,
+      Op.getDebugLoc(),
+      MVT::i1,
+      Op, DAG.getConstantFP(0.0f, MVT::f32),
+      DAG.getCondCode(ISD::SETNE)
+      );
+}
+
+SDValue R600TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
+  SDValue Chain = Op.getOperand(0);
+  SDValue CC = Op.getOperand(1);
+  SDValue LHS   = Op.getOperand(2);
+  SDValue RHS   = Op.getOperand(3);
+  SDValue JumpT  = Op.getOperand(4);
+  SDValue CmpValue;
+  SDValue Result;
+
+  if (LHS.getValueType() == MVT::i32) {
+    CmpValue = DAG.getNode(
+        ISD::SELECT_CC,
+        Op.getDebugLoc(),
+        MVT::i32,
+        LHS, RHS,
+        DAG.getConstant(-1, MVT::i32),
+        DAG.getConstant(0, MVT::i32),
+        CC);
+  } else if (LHS.getValueType() == MVT::f32) {
+    CmpValue = DAG.getNode(
+        ISD::SELECT_CC,
+        Op.getDebugLoc(),
+        MVT::f32,
+        LHS, RHS,
+        DAG.getConstantFP(1.0f, MVT::f32),
+        DAG.getConstantFP(0.0f, MVT::f32),
+        CC);
+  } else {
+    assert(0 && "Not valid type for br_cc");
+  }
+  Result = DAG.getNode(
+      AMDGPUISD::BRANCH_COND,
+      CmpValue.getDebugLoc(),
+      MVT::Other, Chain,
+      JumpT, CmpValue);
+  return Result;
+}
+
+SDValue R600TargetLowering::LowerImplicitParameter(SelectionDAG &DAG, EVT VT,
+                                                   DebugLoc DL,
+                                                   unsigned DwordOffset) const {
+  unsigned ByteOffset = DwordOffset * 4;
+  PointerType * PtrType = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
+                                      AMDGPUAS::PARAM_I_ADDRESS);
+
+  // We shouldn't be using an offset wider than 16-bits for implicit parameters.
+  assert(isInt<16>(ByteOffset));
+
+  return DAG.getLoad(VT, DL, DAG.getEntryNode(),
+                     DAG.getConstant(ByteOffset, MVT::i32), // PTR
+                     MachinePointerInfo(ConstantPointerNull::get(PtrType)),
+                     false, false, false, 0);
+}
+
+SDValue R600TargetLowering::LowerROTL(SDValue Op, SelectionDAG &DAG) const {
+  DebugLoc DL = Op.getDebugLoc();
+  EVT VT = Op.getValueType();
+
+  return DAG.getNode(AMDGPUISD::BITALIGN, DL, VT,
+                     Op.getOperand(0),
+                     Op.getOperand(0),
+                     DAG.getNode(ISD::SUB, DL, VT,
+                                 DAG.getConstant(32, MVT::i32),
+                                 Op.getOperand(1)));
+}
+
+bool R600TargetLowering::isZero(SDValue Op) const {
+  if(ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op)) {
+    return Cst->isNullValue();
+  } else if(ConstantFPSDNode *CstFP = dyn_cast<ConstantFPSDNode>(Op)){
+    return CstFP->isZero();
+  } else {
+    return false;
+  }
+}
+
+SDValue R600TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
+  DebugLoc DL = Op.getDebugLoc();
+  EVT VT = Op.getValueType();
+
+  SDValue LHS = Op.getOperand(0);
+  SDValue RHS = Op.getOperand(1);
+  SDValue True = Op.getOperand(2);
+  SDValue False = Op.getOperand(3);
+  SDValue CC = Op.getOperand(4);
+  SDValue Temp;
+
+  // LHS and RHS are guaranteed to be the same value type
+  EVT CompareVT = LHS.getValueType();
+
+  // Check if we can lower this to a native operation.
+
+  // Try to lower to a CND* instruction:
+  // CND* instructions requires RHS to be zero.  Some SELECT_CC nodes that
+  // can be lowered to CND* instructions can also be lowered to SET*
+  // instructions.  CND* instructions are cheaper, because they dont't
+  // require additional instructions to convert their result to the correct
+  // value type, so this check should be first.
+  if (isZero(LHS) || isZero(RHS)) {
+    SDValue Cond = (isZero(LHS) ? RHS : LHS);
+    SDValue Zero = (isZero(LHS) ? LHS : RHS);
+    ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
+    if (CompareVT != VT) {
+      // Bitcast True / False to the correct types.  This will end up being
+      // a nop, but it allows us to define only a single pattern in the
+      // .TD files for each CND* instruction rather than having to have
+      // one pattern for integer True/False and one for fp True/False
+      True = DAG.getNode(ISD::BITCAST, DL, CompareVT, True);
+      False = DAG.getNode(ISD::BITCAST, DL, CompareVT, False);
+    }
+    if (isZero(LHS)) {
+      CCOpcode = ISD::getSetCCSwappedOperands(CCOpcode);
+    }
+
+    switch (CCOpcode) {
+    case ISD::SETONE:
+    case ISD::SETUNE:
+    case ISD::SETNE:
+    case ISD::SETULE:
+    case ISD::SETULT:
+    case ISD::SETOLE:
+    case ISD::SETOLT:
+    case ISD::SETLE:
+    case ISD::SETLT:
+      CCOpcode = ISD::getSetCCInverse(CCOpcode, CompareVT == MVT::i32);
+      Temp = True;
+      True = False;
+      False = Temp;
+      break;
+    default:
+      break;
+    }
+    SDValue SelectNode = DAG.getNode(ISD::SELECT_CC, DL, CompareVT,
+        Cond, Zero,
+        True, False,
+        DAG.getCondCode(CCOpcode));
+    return DAG.getNode(ISD::BITCAST, DL, VT, SelectNode);
+  }
+
+  // Try to lower to a SET* instruction:
+  // We need all the operands of SELECT_CC to have the same value type, so if
+  // necessary we need to change True and False to be the same type as LHS and
+  // RHS, and then convert the result of the select_cc back to the correct type.
+
+  // Move hardware True/False values to the correct operand.
+  if (isHWTrueValue(False) && isHWFalseValue(True)) {
+    ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
+    std::swap(False, True);
+    CC = DAG.getCondCode(ISD::getSetCCInverse(CCOpcode, CompareVT == MVT::i32));
+  }
+
+  if (isHWTrueValue(True) && isHWFalseValue(False)) {
+    if (CompareVT !=  VT) {
+      if (VT == MVT::f32 && CompareVT == MVT::i32) {
+        SDValue Boolean = DAG.getNode(ISD::SELECT_CC, DL, CompareVT,
+            LHS, RHS,
+            DAG.getConstant(-1, MVT::i32),
+            DAG.getConstant(0, MVT::i32),
+            CC);
+        // Convert integer values of true (-1) and false (0) to fp values of
+        // true (1.0f) and false (0.0f).
+        SDValue LSB = DAG.getNode(ISD::AND, DL, MVT::i32, Boolean,
+                                                  DAG.getConstant(1, MVT::i32));
+        return DAG.getNode(ISD::UINT_TO_FP, DL, VT, LSB);
+      } else if (VT == MVT::i32 && CompareVT == MVT::f32) {
+        SDValue BoolAsFlt = DAG.getNode(ISD::SELECT_CC, DL, CompareVT,
+            LHS, RHS,
+            DAG.getConstantFP(1.0f, MVT::f32),
+            DAG.getConstantFP(0.0f, MVT::f32),
+            CC);
+        // Convert fp values of true (1.0f) and false (0.0f) to integer values
+        // of true (-1) and false (0).
+        SDValue Neg = DAG.getNode(ISD::FNEG, DL, MVT::f32, BoolAsFlt);
+        return DAG.getNode(ISD::FP_TO_SINT, DL, VT, Neg);
+      } else {
+        // I don't think there will be any other type pairings.
+        assert(!"Unhandled operand type parings in SELECT_CC");
+      }
+    } else {
+      // This SELECT_CC is already legal.
+      return DAG.getNode(ISD::SELECT_CC, DL, VT, LHS, RHS, True, False, CC);
+    }
+  }
+
+  // Possible Min/Max pattern
+  SDValue MinMax = LowerMinMax(Op, DAG);
+  if (MinMax.getNode()) {
+    return MinMax;
+  }
+
+  // If we make it this for it means we have no native instructions to handle
+  // this SELECT_CC, so we must lower it.
+  SDValue HWTrue, HWFalse;
+
+  if (CompareVT == MVT::f32) {
+    HWTrue = DAG.getConstantFP(1.0f, CompareVT);
+    HWFalse = DAG.getConstantFP(0.0f, CompareVT);
+  } else if (CompareVT == MVT::i32) {
+    HWTrue = DAG.getConstant(-1, CompareVT);
+    HWFalse = DAG.getConstant(0, CompareVT);
+  }
+  else {
+    assert(!"Unhandled value type in LowerSELECT_CC");
+  }
+
+  // Lower this unsupported SELECT_CC into a combination of two supported
+  // SELECT_CC operations.
+  SDValue Cond = DAG.getNode(ISD::SELECT_CC, DL, CompareVT, LHS, RHS, HWTrue, HWFalse, CC);
+
+  return DAG.getNode(ISD::SELECT_CC, DL, VT,
+      Cond, HWFalse,
+      True, False,
+      DAG.getCondCode(ISD::SETNE));
+}
+
+SDValue R600TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
+  return DAG.getNode(ISD::SELECT_CC,
+      Op.getDebugLoc(),
+      Op.getValueType(),
+      Op.getOperand(0),
+      DAG.getConstant(0, MVT::i32),
+      Op.getOperand(1),
+      Op.getOperand(2),
+      DAG.getCondCode(ISD::SETNE));
+}
+
+SDValue R600TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
+  SDValue Cond;
+  SDValue LHS = Op.getOperand(0);
+  SDValue RHS = Op.getOperand(1);
+  SDValue CC  = Op.getOperand(2);
+  DebugLoc DL = Op.getDebugLoc();
+  assert(Op.getValueType() == MVT::i32);
+  if (LHS.getValueType() == MVT::i32) {
+    Cond = DAG.getNode(
+        ISD::SELECT_CC,
+        Op.getDebugLoc(),
+        MVT::i32,
+        LHS, RHS,
+        DAG.getConstant(-1, MVT::i32),
+        DAG.getConstant(0, MVT::i32),
+        CC);
+  } else if (LHS.getValueType() == MVT::f32) {
+    Cond = DAG.getNode(
+        ISD::SELECT_CC,
+        Op.getDebugLoc(),
+        MVT::f32,
+        LHS, RHS,
+        DAG.getConstantFP(1.0f, MVT::f32),
+        DAG.getConstantFP(0.0f, MVT::f32),
+        CC);
+    Cond = DAG.getNode(
+        ISD::FP_TO_SINT,
+        DL,
+        MVT::i32,
+        Cond);
+  } else {
+    assert(0 && "Not valid type for set_cc");
+  }
+  Cond = DAG.getNode(
+      ISD::AND,
+      DL,
+      MVT::i32,
+      DAG.getConstant(1, MVT::i32),
+      Cond);
+  return Cond;
+}
+
+SDValue R600TargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
+  DebugLoc DL = Op.getDebugLoc();
+  StoreSDNode *StoreNode = cast<StoreSDNode>(Op);
+  SDValue Chain = Op.getOperand(0);
+  SDValue Value = Op.getOperand(1);
+  SDValue Ptr = Op.getOperand(2);
+
+  if (StoreNode->getAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS &&
+      Ptr->getOpcode() != AMDGPUISD::DWORDADDR) {
+    // Convert pointer from byte address to dword address.
+    Ptr = DAG.getNode(AMDGPUISD::DWORDADDR, DL, Ptr.getValueType(),
+                      DAG.getNode(ISD::SRL, DL, Ptr.getValueType(),
+                                  Ptr, DAG.getConstant(2, MVT::i32)));
+
+    if (StoreNode->isTruncatingStore() || StoreNode->isIndexed()) {
+      assert(!"Truncated and indexed stores not supported yet");
+    } else {
+      Chain = DAG.getStore(Chain, DL, Value, Ptr, StoreNode->getMemOperand());
+    }
+    return Chain;
+  }
+  return SDValue();
+}
+
+
+SDValue R600TargetLowering::LowerFPOW(SDValue Op,
+    SelectionDAG &DAG) const {
+  DebugLoc DL = Op.getDebugLoc();
+  EVT VT = Op.getValueType();
+  SDValue LogBase = DAG.getNode(ISD::FLOG2, DL, VT, Op.getOperand(0));
+  SDValue MulLogBase = DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), LogBase);
+  return DAG.getNode(ISD::FEXP2, DL, VT, MulLogBase);
+}
+
+/// XXX Only kernel functions are supported, so we can assume for now that
+/// every function is a kernel function, but in the future we should use
+/// separate calling conventions for kernel and non-kernel functions.
+SDValue R600TargetLowering::LowerFormalArguments(
+                                      SDValue Chain,
+                                      CallingConv::ID CallConv,
+                                      bool isVarArg,
+                                      const SmallVectorImpl<ISD::InputArg> &Ins,
+                                      DebugLoc DL, SelectionDAG &DAG,
+                                      SmallVectorImpl<SDValue> &InVals) const {
+  unsigned ParamOffsetBytes = 36;
+  Function::const_arg_iterator FuncArg =
+                            DAG.getMachineFunction().getFunction()->arg_begin();
+  for (unsigned i = 0, e = Ins.size(); i < e; ++i, ++FuncArg) {
+    EVT VT = Ins[i].VT;
+    Type *ArgType = FuncArg->getType();
+    unsigned ArgSizeInBits = ArgType->isPointerTy() ?
+                             32 : ArgType->getPrimitiveSizeInBits();
+    unsigned ArgBytes = ArgSizeInBits >> 3;
+    EVT ArgVT;
+    if (ArgSizeInBits < VT.getSizeInBits()) {
+      assert(!ArgType->isFloatTy() &&
+             "Extending floating point arguments not supported yet");
+      ArgVT = MVT::getIntegerVT(ArgSizeInBits);
+    } else {
+      ArgVT = VT;
+    }
+    PointerType *PtrTy = PointerType::get(VT.getTypeForEVT(*DAG.getContext()),
+                                                    AMDGPUAS::PARAM_I_ADDRESS);
+    SDValue Arg = DAG.getExtLoad(ISD::ZEXTLOAD, DL, VT, DAG.getRoot(),
+                                DAG.getConstant(ParamOffsetBytes, MVT::i32),
+                                       MachinePointerInfo(new Argument(PtrTy)),
+                                       ArgVT, false, false, ArgBytes);
+    InVals.push_back(Arg);
+    ParamOffsetBytes += ArgBytes;
+  }
+  return Chain;
+}
+
+EVT R600TargetLowering::getSetCCResultType(EVT VT) const {
+   if (!VT.isVector()) return MVT::i32;
+   return VT.changeVectorElementTypeToInteger();
+}
+
+//===----------------------------------------------------------------------===//
+// Custom DAG Optimizations
+//===----------------------------------------------------------------------===//
+
+SDValue R600TargetLowering::PerformDAGCombine(SDNode *N,
+                                              DAGCombinerInfo &DCI) const {
+  SelectionDAG &DAG = DCI.DAG;
+
+  switch (N->getOpcode()) {
+  // (f32 fp_round (f64 uint_to_fp a)) -> (f32 uint_to_fp a)
+  case ISD::FP_ROUND: {
+      SDValue Arg = N->getOperand(0);
+      if (Arg.getOpcode() == ISD::UINT_TO_FP && Arg.getValueType() == MVT::f64) {
+        return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), N->getValueType(0),
+                           Arg.getOperand(0));
+      }
+      break;
+    }
+  }
+  return SDValue();
+}
diff --git a/lib/Target/R600/R600ISelLowering.h b/lib/Target/R600/R600ISelLowering.h
new file mode 100644
index 0000000000..2b954dab55
--- /dev/null
+++ b/lib/Target/R600/R600ISelLowering.h
@@ -0,0 +1,72 @@
+//===-- R600ISelLowering.h - R600 DAG Lowering Interface -*- C++ -*--------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief R600 DAG Lowering interface definition
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef R600ISELLOWERING_H
+#define R600ISELLOWERING_H
+
+#include "AMDGPUISelLowering.h"
+
+namespace llvm {
+
+class R600InstrInfo;
+
+class R600TargetLowering : public AMDGPUTargetLowering {
+public:
+  R600TargetLowering(TargetMachine &TM);
+  virtual MachineBasicBlock * EmitInstrWithCustomInserter(MachineInstr *MI,
+      MachineBasicBlock * BB) const;
+  virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
+  virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
+  void ReplaceNodeResults(SDNode * N,
+      SmallVectorImpl<SDValue> &Results,
+      SelectionDAG &DAG) const;
+  virtual SDValue LowerFormalArguments(
+                                      SDValue Chain,
+                                      CallingConv::ID CallConv,
+                                      bool isVarArg,
+                                      const SmallVectorImpl<ISD::InputArg> &Ins,
+                                      DebugLoc DL, SelectionDAG &DAG,
+                                      SmallVectorImpl<SDValue> &InVals) const;
+  virtual EVT getSetCCResultType(EVT VT) const;
+private:
+  const R600InstrInfo * TII;
+
+  /// Each OpenCL kernel has nine implicit parameters that are stored in the
+  /// first nine dwords of a Vertex Buffer.  These implicit parameters are
+  /// lowered to load instructions which retreive the values from the Vertex
+  /// Buffer.
+  SDValue LowerImplicitParameter(SelectionDAG &DAG, EVT VT,
+                                 DebugLoc DL, unsigned DwordOffset) const;
+
+  void lowerImplicitParameter(MachineInstr *MI, MachineBasicBlock &BB,
+      MachineRegisterInfo & MRI, unsigned dword_offset) const;
+
+  SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
+
+  /// \brief Lower ROTL opcode to BITALIGN
+  SDValue LowerROTL(SDValue Op, SelectionDAG &DAG) const;
+
+  SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerSELECT(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerSETCC(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerSTORE(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerFPTOUINT(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerFPOW(SDValue Op, SelectionDAG &DAG) const;
+  
+  bool isZero(SDValue Op) const;
+};
+
+} // End namespace llvm;
+
+#endif // R600ISELLOWERING_H
diff --git a/lib/Target/R600/R600InstrInfo.cpp b/lib/Target/R600/R600InstrInfo.cpp
new file mode 100644
index 0000000000..a60a1802d4
--- /dev/null
+++ b/lib/Target/R600/R600InstrInfo.cpp
@@ -0,0 +1,663 @@
+//===-- R600InstrInfo.cpp - R600 Instruction Information ------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief R600 Implementation of TargetInstrInfo.
+//
+//===----------------------------------------------------------------------===//
+
+#include "R600InstrInfo.h"
+#include "AMDGPUTargetMachine.h"
+#include "AMDGPUSubtarget.h"
+#include "R600Defines.h"
+#include "R600RegisterInfo.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+
+#define GET_INSTRINFO_CTOR
+#include "AMDGPUGenDFAPacketizer.inc"
+
+using namespace llvm;
+
+R600InstrInfo::R600InstrInfo(AMDGPUTargetMachine &tm)
+  : AMDGPUInstrInfo(tm),
+    RI(tm, *this)
+  { }
+
+const R600RegisterInfo &R600InstrInfo::getRegisterInfo() const {
+  return RI;
+}
+
+bool R600InstrInfo::isTrig(const MachineInstr &MI) const {
+  return get(MI.getOpcode()).TSFlags & R600_InstFlag::TRIG;
+}
+
+bool R600InstrInfo::isVector(const MachineInstr &MI) const {
+  return get(MI.getOpcode()).TSFlags & R600_InstFlag::VECTOR;
+}
+
+void
+R600InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
+                           MachineBasicBlock::iterator MI, DebugLoc DL,
+                           unsigned DestReg, unsigned SrcReg,
+                           bool KillSrc) const {
+  if (AMDGPU::R600_Reg128RegClass.contains(DestReg)
+      && AMDGPU::R600_Reg128RegClass.contains(SrcReg)) {
+    for (unsigned I = 0; I < 4; I++) {
+      unsigned SubRegIndex = RI.getSubRegFromChannel(I);
+      buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
+                              RI.getSubReg(DestReg, SubRegIndex),
+                              RI.getSubReg(SrcReg, SubRegIndex))
+                              .addReg(DestReg,
+                                      RegState::Define | RegState::Implicit);
+    }
+  } else {
+
+    // We can't copy vec4 registers
+    assert(!AMDGPU::R600_Reg128RegClass.contains(DestReg)
+           && !AMDGPU::R600_Reg128RegClass.contains(SrcReg));
+
+    MachineInstr *NewMI = buildDefaultInstruction(MBB, MI, AMDGPU::MOV,
+                                                  DestReg, SrcReg);
+    NewMI->getOperand(getOperandIdx(*NewMI, R600Operands::SRC0))
+                                    .setIsKill(KillSrc);
+  }
+}
+
+MachineInstr * R600InstrInfo::getMovImmInstr(MachineFunction *MF,
+                                             unsigned DstReg, int64_t Imm) const {
+  MachineInstr * MI = MF->CreateMachineInstr(get(AMDGPU::MOV), DebugLoc());
+  MachineInstrBuilder(MI).addReg(DstReg, RegState::Define);
+  MachineInstrBuilder(MI).addReg(AMDGPU::ALU_LITERAL_X);
+  MachineInstrBuilder(MI).addImm(Imm);
+  MachineInstrBuilder(MI).addReg(0); // PREDICATE_BIT
+
+  return MI;
+}
+
+unsigned R600InstrInfo::getIEQOpcode() const {
+  return AMDGPU::SETE_INT;
+}
+
+bool R600InstrInfo::isMov(unsigned Opcode) const {
+
+
+  switch(Opcode) {
+  default: return false;
+  case AMDGPU::MOV:
+  case AMDGPU::MOV_IMM_F32:
+  case AMDGPU::MOV_IMM_I32:
+    return true;
+  }
+}
+
+// Some instructions act as place holders to emulate operations that the GPU
+// hardware does automatically. This function can be used to check if
+// an opcode falls into this category.
+bool R600InstrInfo::isPlaceHolderOpcode(unsigned Opcode) const {
+  switch (Opcode) {
+  default: return false;
+  case AMDGPU::RETURN:
+  case AMDGPU::RESERVE_REG:
+    return true;
+  }
+}
+
+bool R600InstrInfo::isReductionOp(unsigned Opcode) const {
+  switch(Opcode) {
+    default: return false;
+    case AMDGPU::DOT4_r600_pseudo:
+    case AMDGPU::DOT4_eg_pseudo:
+      return true;
+  }
+}
+
+bool R600InstrInfo::isCubeOp(unsigned Opcode) const {
+  switch(Opcode) {
+    default: return false;
+    case AMDGPU::CUBE_r600_pseudo:
+    case AMDGPU::CUBE_r600_real:
+    case AMDGPU::CUBE_eg_pseudo:
+    case AMDGPU::CUBE_eg_real:
+      return true;
+  }
+}
+
+bool R600InstrInfo::isALUInstr(unsigned Opcode) const {
+  unsigned TargetFlags = get(Opcode).TSFlags;
+
+  return ((TargetFlags & R600_InstFlag::OP1) |
+          (TargetFlags & R600_InstFlag::OP2) |
+          (TargetFlags & R600_InstFlag::OP3));
+}
+
+DFAPacketizer *R600InstrInfo::CreateTargetScheduleState(const TargetMachine *TM,
+    const ScheduleDAG *DAG) const {
+  const InstrItineraryData *II = TM->getInstrItineraryData();
+  return TM->getSubtarget<AMDGPUSubtarget>().createDFAPacketizer(II);
+}
+
+static bool
+isPredicateSetter(unsigned Opcode) {
+  switch (Opcode) {
+  case AMDGPU::PRED_X:
+    return true;
+  default:
+    return false;
+  }
+}
+
+static MachineInstr *
+findFirstPredicateSetterFrom(MachineBasicBlock &MBB,
+                             MachineBasicBlock::iterator I) {
+  while (I != MBB.begin()) {
+    --I;
+    MachineInstr *MI = I;
+    if (isPredicateSetter(MI->getOpcode()))
+      return MI;
+  }
+
+  return NULL;
+}
+
+bool
+R600InstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,
+                             MachineBasicBlock *&TBB,
+                             MachineBasicBlock *&FBB,
+                             SmallVectorImpl<MachineOperand> &Cond,
+                             bool AllowModify) const {
+  // Most of the following comes from the ARM implementation of AnalyzeBranch
+
+  // If the block has no terminators, it just falls into the block after it.
+  MachineBasicBlock::iterator I = MBB.end();
+  if (I == MBB.begin())
+    return false;
+  --I;
+  while (I->isDebugValue()) {
+    if (I == MBB.begin())
+      return false;
+    --I;
+  }
+  if (static_cast<MachineInstr *>(I)->getOpcode() != AMDGPU::JUMP) {
+    return false;
+  }
+
+  // Get the last instruction in the block.
+  MachineInstr *LastInst = I;
+
+  // If there is only one terminator instruction, process it.
+  unsigned LastOpc = LastInst->getOpcode();
+  if (I == MBB.begin() ||
+      static_cast<MachineInstr *>(--I)->getOpcode() != AMDGPU::JUMP) {
+    if (LastOpc == AMDGPU::JUMP) {
+      if(!isPredicated(LastInst)) {
+        TBB = LastInst->getOperand(0).getMBB();
+        return false;
+      } else {
+        MachineInstr *predSet = I;
+        while (!isPredicateSetter(predSet->getOpcode())) {
+          predSet = --I;
+        }
+        TBB = LastInst->getOperand(0).getMBB();
+        Cond.push_back(predSet->getOperand(1));
+        Cond.push_back(predSet->getOperand(2));
+        Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
+        return false;
+      }
+    }
+    return true;  // Can't handle indirect branch.
+  }
+
+  // Get the instruction before it if it is a terminator.
+  MachineInstr *SecondLastInst = I;
+  unsigned SecondLastOpc = SecondLastInst->getOpcode();
+
+  // If the block ends with a B and a Bcc, handle it.
+  if (SecondLastOpc == AMDGPU::JUMP &&
+      isPredicated(SecondLastInst) &&
+      LastOpc == AMDGPU::JUMP &&
+      !isPredicated(LastInst)) {
+    MachineInstr *predSet = --I;
+    while (!isPredicateSetter(predSet->getOpcode())) {
+      predSet = --I;
+    }
+    TBB = SecondLastInst->getOperand(0).getMBB();
+    FBB = LastInst->getOperand(0).getMBB();
+    Cond.push_back(predSet->getOperand(1));
+    Cond.push_back(predSet->getOperand(2));
+    Cond.push_back(MachineOperand::CreateReg(AMDGPU::PRED_SEL_ONE, false));
+    return false;
+  }
+
+  // Otherwise, can't handle this.
+  return true;
+}
+
+int R600InstrInfo::getBranchInstr(const MachineOperand &op) const {
+  const MachineInstr *MI = op.getParent();
+
+  switch (MI->getDesc().OpInfo->RegClass) {
+  default: // FIXME: fallthrough??
+  case AMDGPU::GPRI32RegClassID: return AMDGPU::BRANCH_COND_i32;
+  case AMDGPU::GPRF32RegClassID: return AMDGPU::BRANCH_COND_f32;
+  };
+}
+
+unsigned
+R600InstrInfo::InsertBranch(MachineBasicBlock &MBB,
+                            MachineBasicBlock *TBB,
+                            MachineBasicBlock *FBB,
+                            const SmallVectorImpl<MachineOperand> &Cond,
+                            DebugLoc DL) const {
+  assert(TBB && "InsertBranch must not be told to insert a fallthrough");
+
+  if (FBB == 0) {
+    if (Cond.empty()) {
+      BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(TBB).addReg(0);
+      return 1;
+    } else {
+      MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
+      assert(PredSet && "No previous predicate !");
+      addFlag(PredSet, 0, MO_FLAG_PUSH);
+      PredSet->getOperand(2).setImm(Cond[1].getImm());
+
+      BuildMI(&MBB, DL, get(AMDGPU::JUMP))
+             .addMBB(TBB)
+             .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
+      return 1;
+    }
+  } else {
+    MachineInstr *PredSet = findFirstPredicateSetterFrom(MBB, MBB.end());
+    assert(PredSet && "No previous predicate !");
+    addFlag(PredSet, 0, MO_FLAG_PUSH);
+    PredSet->getOperand(2).setImm(Cond[1].getImm());
+    BuildMI(&MBB, DL, get(AMDGPU::JUMP))
+            .addMBB(TBB)
+            .addReg(AMDGPU::PREDICATE_BIT, RegState::Kill);
+    BuildMI(&MBB, DL, get(AMDGPU::JUMP)).addMBB(FBB).addReg(0);
+    return 2;
+  }
+}
+
+unsigned
+R600InstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
+
+  // Note : we leave PRED* instructions there.
+  // They may be needed when predicating instructions.
+
+  MachineBasicBlock::iterator I = MBB.end();
+
+  if (I == MBB.begin()) {
+    return 0;
+  }
+  --I;
+  switch (I->getOpcode()) {
+  default:
+    return 0;
+  case AMDGPU::JUMP:
+    if (isPredicated(I)) {
+      MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
+      clearFlag(predSet, 0, MO_FLAG_PUSH);
+    }
+    I->eraseFromParent();
+    break;
+  }
+  I = MBB.end();
+
+  if (I == MBB.begin()) {
+    return 1;
+  }
+  --I;
+  switch (I->getOpcode()) {
+    // FIXME: only one case??
+  default:
+    return 1;
+  case AMDGPU::JUMP:
+    if (isPredicated(I)) {
+      MachineInstr *predSet = findFirstPredicateSetterFrom(MBB, I);
+      clearFlag(predSet, 0, MO_FLAG_PUSH);
+    }
+    I->eraseFromParent();
+    break;
+  }
+  return 2;
+}
+
+bool
+R600InstrInfo::isPredicated(const MachineInstr *MI) const {
+  int idx = MI->findFirstPredOperandIdx();
+  if (idx < 0)
+    return false;
+
+  unsigned Reg = MI->getOperand(idx).getReg();
+  switch (Reg) {
+  default: return false;
+  case AMDGPU::PRED_SEL_ONE:
+  case AMDGPU::PRED_SEL_ZERO:
+  case AMDGPU::PREDICATE_BIT:
+    return true;
+  }
+}
+
+bool
+R600InstrInfo::isPredicable(MachineInstr *MI) const {
+  // XXX: KILL* instructions can be predicated, but they must be the last
+  // instruction in a clause, so this means any instructions after them cannot
+  // be predicated.  Until we have proper support for instruction clauses in the
+  // backend, we will mark KILL* instructions as unpredicable.
+
+  if (MI->getOpcode() == AMDGPU::KILLGT) {
+    return false;
+  } else {
+    return AMDGPUInstrInfo::isPredicable(MI);
+  }
+}
+
+
+bool
+R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
+                                   unsigned NumCyles,
+                                   unsigned ExtraPredCycles,
+                                   const BranchProbability &Probability) const{
+  return true;
+}
+
+bool
+R600InstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
+                                   unsigned NumTCycles,
+                                   unsigned ExtraTCycles,
+                                   MachineBasicBlock &FMBB,
+                                   unsigned NumFCycles,
+                                   unsigned ExtraFCycles,
+                                   const BranchProbability &Probability) const {
+  return true;
+}
+
+bool
+R600InstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
+                                         unsigned NumCyles,
+                                         const BranchProbability &Probability)
+                                         const {
+  return true;
+}
+
+bool
+R600InstrInfo::isProfitableToUnpredicate(MachineBasicBlock &TMBB,
+                                         MachineBasicBlock &FMBB) const {
+  return false;
+}
+
+
+bool
+R600InstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const {
+  MachineOperand &MO = Cond[1];
+  switch (MO.getImm()) {
+  case OPCODE_IS_ZERO_INT:
+    MO.setImm(OPCODE_IS_NOT_ZERO_INT);
+    break;
+  case OPCODE_IS_NOT_ZERO_INT:
+    MO.setImm(OPCODE_IS_ZERO_INT);
+    break;
+  case OPCODE_IS_ZERO:
+    MO.setImm(OPCODE_IS_NOT_ZERO);
+    break;
+  case OPCODE_IS_NOT_ZERO:
+    MO.setImm(OPCODE_IS_ZERO);
+    break;
+  default:
+    return true;
+  }
+
+  MachineOperand &MO2 = Cond[2];
+  switch (MO2.getReg()) {
+  case AMDGPU::PRED_SEL_ZERO:
+    MO2.setReg(AMDGPU::PRED_SEL_ONE);
+    break;
+  case AMDGPU::PRED_SEL_ONE:
+    MO2.setReg(AMDGPU::PRED_SEL_ZERO);
+    break;
+  default:
+    return true;
+  }
+  return false;
+}
+
+bool
+R600InstrInfo::DefinesPredicate(MachineInstr *MI,
+                                std::vector<MachineOperand> &Pred) const {
+  return isPredicateSetter(MI->getOpcode());
+}
+
+
+bool
+R600InstrInfo::SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
+                       const SmallVectorImpl<MachineOperand> &Pred2) const {
+  return false;
+}
+
+
+bool
+R600InstrInfo::PredicateInstruction(MachineInstr *MI,
+                      const SmallVectorImpl<MachineOperand> &Pred) const {
+  int PIdx = MI->findFirstPredOperandIdx();
+
+  if (PIdx != -1) {
+    MachineOperand &PMO = MI->getOperand(PIdx);
+    PMO.setReg(Pred[2].getReg());
+    MachineInstrBuilder(MI).addReg(AMDGPU::PREDICATE_BIT, RegState::Implicit);
+    return true;
+  }
+
+  return false;
+}
+
+unsigned int R600InstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
+                                            const MachineInstr *MI,
+                                            unsigned *PredCost) const {
+  if (PredCost)
+    *PredCost = 2;
+  return 2;
+}
+
+MachineInstrBuilder R600InstrInfo::buildDefaultInstruction(MachineBasicBlock &MBB,
+                                                  MachineBasicBlock::iterator I,
+                                                  unsigned Opcode,
+                                                  unsigned DstReg,
+                                                  unsigned Src0Reg,
+                                                  unsigned Src1Reg) const {
+  MachineInstrBuilder MIB = BuildMI(MBB, I, MBB.findDebugLoc(I), get(Opcode),
+    DstReg);           // $dst
+
+  if (Src1Reg) {
+    MIB.addImm(0)     // $update_exec_mask
+       .addImm(0);    // $update_predicate
+  }
+  MIB.addImm(1)        // $write
+     .addImm(0)        // $omod
+     .addImm(0)        // $dst_rel
+     .addImm(0)        // $dst_clamp
+     .addReg(Src0Reg)  // $src0
+     .addImm(0)        // $src0_neg
+     .addImm(0)        // $src0_rel
+     .addImm(0);       // $src0_abs
+
+  if (Src1Reg) {
+    MIB.addReg(Src1Reg) // $src1
+       .addImm(0)       // $src1_neg
+       .addImm(0)       // $src1_rel
+       .addImm(0);       // $src1_abs
+  }
+
+  //XXX: The r600g finalizer expects this to be 1, once we've moved the
+  //scheduling to the backend, we can change the default to 0.
+  MIB.addImm(1)        // $last
+      .addReg(AMDGPU::PRED_SEL_OFF) // $pred_sel
+      .addImm(0);        // $literal
+
+  return MIB;
+}
+
+MachineInstr *R600InstrInfo::buildMovImm(MachineBasicBlock &BB,
+                                         MachineBasicBlock::iterator I,
+                                         unsigned DstReg,
+                                         uint64_t Imm) const {
+  MachineInstr *MovImm = buildDefaultInstruction(BB, I, AMDGPU::MOV, DstReg,
+                                                  AMDGPU::ALU_LITERAL_X);
+  setImmOperand(MovImm, R600Operands::IMM, Imm);
+  return MovImm;
+}
+
+int R600InstrInfo::getOperandIdx(const MachineInstr &MI,
+                                 R600Operands::Ops Op) const {
+  return getOperandIdx(MI.getOpcode(), Op);
+}
+
+int R600InstrInfo::getOperandIdx(unsigned Opcode,
+                                 R600Operands::Ops Op) const {
+  const static int OpTable[3][R600Operands::COUNT] = {
+//            W        C     S  S  S     S  S  S     S  S
+//            R  O  D  L  S  R  R  R  S  R  R  R  S  R  R  L  P
+//   D  U     I  M  R  A  R  C  C  C  C  C  C  C  R  C  C  A  R  I
+//   S  E  U  T  O  E  M  C  0  0  0  C  1  1  1  C  2  2  S  E  M
+//   T  M  P  E  D  L  P  0  N  R  A  1  N  R  A  2  N  R  T  D  M
+    {0,-1,-1, 1, 2, 3, 4, 5, 6, 7, 8,-1,-1,-1,-1,-1,-1,-1, 9,10,11},
+    {0, 1, 2, 3, 4 ,5 ,6 ,7, 8, 9,10,11,12,-1,-1,-1,13,14,15,16,17},
+    {0,-1,-1,-1,-1, 1, 2, 3, 4, 5,-1, 6, 7, 8,-1, 9,10,11,12,13,14}
+  };
+  unsigned TargetFlags = get(Opcode).TSFlags;
+  unsigned OpTableIdx;
+
+  if (!HAS_NATIVE_OPERANDS(TargetFlags)) {
+    switch (Op) {
+    case R600Operands::DST: return 0;
+    case R600Operands::SRC0: return 1;
+    case R600Operands::SRC1: return 2;
+    case R600Operands::SRC2: return 3;
+    default:
+      assert(!"Unknown operand type for instruction");
+      return -1;
+    }
+  }
+
+  if (TargetFlags & R600_InstFlag::OP1) {
+    OpTableIdx = 0;
+  } else if (TargetFlags & R600_InstFlag::OP2) {
+    OpTableIdx = 1;
+  } else {
+    assert((TargetFlags & R600_InstFlag::OP3) && "OP1, OP2, or OP3 not defined "
+                                                 "for this instruction");
+    OpTableIdx = 2;
+  }
+
+  return OpTable[OpTableIdx][Op];
+}
+
+void R600InstrInfo::setImmOperand(MachineInstr *MI, R600Operands::Ops Op,
+                                  int64_t Imm) const {
+  int Idx = getOperandIdx(*MI, Op);
+  assert(Idx != -1 && "Operand not supported for this instruction.");
+  assert(MI->getOperand(Idx).isImm());
+  MI->getOperand(Idx).setImm(Imm);
+}
+
+//===----------------------------------------------------------------------===//
+// Instruction flag getters/setters
+//===----------------------------------------------------------------------===//
+
+bool R600InstrInfo::hasFlagOperand(const MachineInstr &MI) const {
+  return GET_FLAG_OPERAND_IDX(get(MI.getOpcode()).TSFlags) != 0;
+}
+
+MachineOperand &R600InstrInfo::getFlagOp(MachineInstr *MI, unsigned SrcIdx,
+                                         unsigned Flag) const {
+  unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
+  int FlagIndex = 0;
+  if (Flag != 0) {
+    // If we pass something other than the default value of Flag to this
+    // function, it means we are want to set a flag on an instruction
+    // that uses native encoding.
+    assert(HAS_NATIVE_OPERANDS(TargetFlags));
+    bool IsOP3 = (TargetFlags & R600_InstFlag::OP3) == R600_InstFlag::OP3;
+    switch (Flag) {
+    case MO_FLAG_CLAMP:
+      FlagIndex = getOperandIdx(*MI, R600Operands::CLAMP);
+      break;
+    case MO_FLAG_MASK:
+      FlagIndex = getOperandIdx(*MI, R600Operands::WRITE);
+      break;
+    case MO_FLAG_NOT_LAST:
+    case MO_FLAG_LAST:
+      FlagIndex = getOperandIdx(*MI, R600Operands::LAST);
+      break;
+    case MO_FLAG_NEG:
+      switch (SrcIdx) {
+      case 0: FlagIndex = getOperandIdx(*MI, R600Operands::SRC0_NEG); break;
+      case 1: FlagIndex = getOperandIdx(*MI, R600Operands::SRC1_NEG); break;
+      case 2: FlagIndex = getOperandIdx(*MI, R600Operands::SRC2_NEG); break;
+      }
+      break;
+
+    case MO_FLAG_ABS:
+      assert(!IsOP3 && "Cannot set absolute value modifier for OP3 "
+                       "instructions.");
+      switch (SrcIdx) {
+      case 0: FlagIndex = getOperandIdx(*MI, R600Operands::SRC0_ABS); break;
+      case 1: FlagIndex = getOperandIdx(*MI, R600Operands::SRC1_ABS); break;
+      }
+      break;
+
+    default:
+      FlagIndex = -1;
+      break;
+    }
+    assert(FlagIndex != -1 && "Flag not supported for this instruction");
+  } else {
+      FlagIndex = GET_FLAG_OPERAND_IDX(TargetFlags);
+      assert(FlagIndex != 0 &&
+         "Instruction flags not supported for this instruction");
+  }
+
+  MachineOperand &FlagOp = MI->getOperand(FlagIndex);
+  assert(FlagOp.isImm());
+  return FlagOp;
+}
+
+void R600InstrInfo::addFlag(MachineInstr *MI, unsigned Operand,
+                            unsigned Flag) const {
+  unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
+  if (Flag == 0) {
+    return;
+  }
+  if (HAS_NATIVE_OPERANDS(TargetFlags)) {
+    MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
+    if (Flag == MO_FLAG_NOT_LAST) {
+      clearFlag(MI, Operand, MO_FLAG_LAST);
+    } else if (Flag == MO_FLAG_MASK) {
+      clearFlag(MI, Operand, Flag);
+    } else {
+      FlagOp.setImm(1);
+    }
+  } else {
+      MachineOperand &FlagOp = getFlagOp(MI, Operand);
+      FlagOp.setImm(FlagOp.getImm() | (Flag << (NUM_MO_FLAGS * Operand)));
+  }
+}
+
+void R600InstrInfo::clearFlag(MachineInstr *MI, unsigned Operand,
+                              unsigned Flag) const {
+  unsigned TargetFlags = get(MI->getOpcode()).TSFlags;
+  if (HAS_NATIVE_OPERANDS(TargetFlags)) {
+    MachineOperand &FlagOp = getFlagOp(MI, Operand, Flag);
+    FlagOp.setImm(0);
+  } else {
+    MachineOperand &FlagOp = getFlagOp(MI);
+    unsigned InstFlags = FlagOp.getImm();
+    InstFlags &= ~(Flag << (NUM_MO_FLAGS * Operand));
+    FlagOp.setImm(InstFlags);
+  }
+}
diff --git a/lib/Target/R600/R600InstrInfo.h b/lib/Target/R600/R600InstrInfo.h
new file mode 100644
index 0000000000..6bb0ca92e4
--- /dev/null
+++ b/lib/Target/R600/R600InstrInfo.h
@@ -0,0 +1,169 @@
+//===-- R600InstrInfo.h - R600 Instruction Info Interface -------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Interface definition for R600InstrInfo
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef R600INSTRUCTIONINFO_H_
+#define R600INSTRUCTIONINFO_H_
+
+#include "AMDIL.h"
+#include "AMDGPUInstrInfo.h"
+#include "R600Defines.h"
+#include "R600RegisterInfo.h"
+
+#include <map>
+
+namespace llvm {
+
+  class AMDGPUTargetMachine;
+  class DFAPacketizer;
+  class ScheduleDAG;
+  class MachineFunction;
+  class MachineInstr;
+  class MachineInstrBuilder;
+
+  class R600InstrInfo : public AMDGPUInstrInfo {
+  private:
+  const R600RegisterInfo RI;
+
+  int getBranchInstr(const MachineOperand &op) const;
+
+  public:
+  explicit R600InstrInfo(AMDGPUTargetMachine &tm);
+
+  const R600RegisterInfo &getRegisterInfo() const;
+  virtual void copyPhysReg(MachineBasicBlock &MBB,
+                           MachineBasicBlock::iterator MI, DebugLoc DL,
+                           unsigned DestReg, unsigned SrcReg,
+                           bool KillSrc) const;
+
+  bool isTrig(const MachineInstr &MI) const;
+  bool isPlaceHolderOpcode(unsigned opcode) const;
+  bool isReductionOp(unsigned opcode) const;
+  bool isCubeOp(unsigned opcode) const;
+
+  /// \returns true if this \p Opcode represents an ALU instruction.
+  bool isALUInstr(unsigned Opcode) const;
+
+  /// \breif Vector instructions are instructions that must fill all
+  /// instruction slots within an instruction group.
+  bool isVector(const MachineInstr &MI) const;
+
+  virtual MachineInstr * getMovImmInstr(MachineFunction *MF, unsigned DstReg,
+                                        int64_t Imm) const;
+
+  virtual unsigned getIEQOpcode() const;
+  virtual bool isMov(unsigned Opcode) const;
+
+  DFAPacketizer *CreateTargetScheduleState(const TargetMachine *TM,
+                                           const ScheduleDAG *DAG) const;
+
+  bool ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const;
+
+  bool AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB,
+                     SmallVectorImpl<MachineOperand> &Cond, bool AllowModify) const;
+
+  unsigned InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, MachineBasicBlock *FBB, const SmallVectorImpl<MachineOperand> &Cond, DebugLoc DL) const;
+
+  unsigned RemoveBranch(MachineBasicBlock &MBB) const;
+
+  bool isPredicated(const MachineInstr *MI) const;
+
+  bool isPredicable(MachineInstr *MI) const;
+
+  bool
+   isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCyles,
+                             const BranchProbability &Probability) const;
+
+  bool isProfitableToIfCvt(MachineBasicBlock &MBB, unsigned NumCyles,
+                           unsigned ExtraPredCycles,
+                           const BranchProbability &Probability) const ;
+
+  bool
+   isProfitableToIfCvt(MachineBasicBlock &TMBB,
+                       unsigned NumTCycles, unsigned ExtraTCycles,
+                       MachineBasicBlock &FMBB,
+                       unsigned NumFCycles, unsigned ExtraFCycles,
+                       const BranchProbability &Probability) const;
+
+  bool DefinesPredicate(MachineInstr *MI,
+                                  std::vector<MachineOperand> &Pred) const;
+
+  bool SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1,
+                         const SmallVectorImpl<MachineOperand> &Pred2) const;
+
+  bool isProfitableToUnpredicate(MachineBasicBlock &TMBB,
+                                          MachineBasicBlock &FMBB) const;
+
+  bool PredicateInstruction(MachineInstr *MI,
+                        const SmallVectorImpl<MachineOperand> &Pred) const;
+
+  unsigned int getInstrLatency(const InstrItineraryData *ItinData,
+                               const MachineInstr *MI,
+                               unsigned *PredCost = 0) const;
+
+  virtual int getInstrLatency(const InstrItineraryData *ItinData,
+                              SDNode *Node) const { return 1;}
+
+  /// You can use this function to avoid manually specifying each instruction
+  /// modifier operand when building a new instruction.
+  ///
+  /// \returns a MachineInstr with all the instruction modifiers initialized
+  /// to their default values.
+  MachineInstrBuilder buildDefaultInstruction(MachineBasicBlock &MBB,
+                                              MachineBasicBlock::iterator I,
+                                              unsigned Opcode,
+                                              unsigned DstReg,
+                                              unsigned Src0Reg,
+                                              unsigned Src1Reg = 0) const;
+
+  MachineInstr *buildMovImm(MachineBasicBlock &BB,
+                                  MachineBasicBlock::iterator I,
+                                  unsigned DstReg,
+                                  uint64_t Imm) const;
+
+  /// \brief Get the index of Op in the MachineInstr.
+  ///
+  /// \returns -1 if the Instruction does not contain the specified \p Op.
+  int getOperandIdx(const MachineInstr &MI, R600Operands::Ops Op) const;
+
+  /// \brief Get the index of \p Op for the given Opcode.
+  ///
+  /// \returns -1 if the Instruction does not contain the specified \p Op.
+  int getOperandIdx(unsigned Opcode, R600Operands::Ops Op) const;
+
+  /// \brief Helper function for setting instruction flag values.
+  void setImmOperand(MachineInstr *MI, R600Operands::Ops Op, int64_t Imm) const;
+
+  /// \returns true if this instruction has an operand for storing target flags.
+  bool hasFlagOperand(const MachineInstr &MI) const;
+
+  ///\brief Add one of the MO_FLAG* flags to the specified \p Operand.
+  void addFlag(MachineInstr *MI, unsigned Operand, unsigned Flag) const;
+
+  ///\brief Determine if the specified \p Flag is set on this \p Operand.
+  bool isFlagSet(const MachineInstr &MI, unsigned Operand, unsigned Flag) const;
+
+  /// \param SrcIdx The register source to set the flag on (e.g src0, src1, src2)
+  /// \param Flag The flag being set.
+  ///
+  /// \returns the operand containing the flags for this instruction.
+  MachineOperand &getFlagOp(MachineInstr *MI, unsigned SrcIdx = 0,
+                            unsigned Flag = 0) const;
+
+  /// \brief Clear the specified flag on the instruction.
+  void clearFlag(MachineInstr *MI, unsigned Operand, unsigned Flag) const;
+};
+
+} // End llvm namespace
+
+#endif // R600INSTRINFO_H_
diff --git a/lib/Target/R600/R600Instructions.td b/lib/Target/R600/R600Instructions.td
new file mode 100644
index 0000000000..105822066c
--- /dev/null
+++ b/lib/Target/R600/R600Instructions.td
@@ -0,0 +1,1659 @@
+//===-- R600Instructions.td - R600 Instruction defs  -------*- tablegen -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// R600 Tablegen instruction definitions
+//
+//===----------------------------------------------------------------------===//
+
+include "R600Intrinsics.td"
+
+class InstR600 <bits<11> inst, dag outs, dag ins, string asm, list<dag> pattern,
+                InstrItinClass itin>
+    : AMDGPUInst <outs, ins, asm, pattern> {
+
+  field bits<64> Inst;
+  bit Trig = 0;
+  bit Op3 = 0;
+  bit isVector = 0;
+  bits<2> FlagOperandIdx = 0;
+  bit Op1 = 0;
+  bit Op2 = 0;
+  bit HasNativeOperands = 0;
+
+  bits<11> op_code = inst;
+  //let Inst = inst;
+  let Namespace = "AMDGPU";
+  let OutOperandList = outs;
+  let InOperandList = ins;
+  let AsmString = asm;
+  let Pattern = pattern;
+  let Itinerary = itin;
+
+  let TSFlags{4} = Trig;
+  let TSFlags{5} = Op3;
+
+  // Vector instructions are instructions that must fill all slots in an
+  // instruction group
+  let TSFlags{6} = isVector;
+  let TSFlags{8-7} = FlagOperandIdx;
+  let TSFlags{9} = HasNativeOperands;
+  let TSFlags{10} = Op1;
+  let TSFlags{11} = Op2;
+}
+
+class InstR600ISA <dag outs, dag ins, string asm, list<dag> pattern> :
+    AMDGPUInst <outs, ins, asm, pattern> {
+  field bits<64> Inst;
+
+  let Namespace = "AMDGPU";
+}
+
+def MEMxi : Operand<iPTR> {
+  let MIOperandInfo = (ops R600_TReg32_X:$ptr, i32imm:$index);
+  let PrintMethod = "printMemOperand";
+}
+
+def MEMrr : Operand<iPTR> {
+  let MIOperandInfo = (ops R600_Reg32:$ptr, R600_Reg32:$index);
+}
+
+// Operands for non-registers
+
+class InstFlag<string PM = "printOperand", int Default = 0>
+    : OperandWithDefaultOps <i32, (ops (i32 Default))> {
+  let PrintMethod = PM;
+}
+
+def LITERAL : InstFlag<"printLiteral">;
+
+def WRITE : InstFlag <"printWrite", 1>;
+def OMOD : InstFlag <"printOMOD">;
+def REL : InstFlag <"printRel">;
+def CLAMP : InstFlag <"printClamp">;
+def NEG : InstFlag <"printNeg">;
+def ABS : InstFlag <"printAbs">;
+def UEM : InstFlag <"printUpdateExecMask">;
+def UP : InstFlag <"printUpdatePred">;
+
+// XXX: The r600g finalizer in Mesa expects last to be one in most cases.
+// Once we start using the packetizer in this backend we should have this
+// default to 0.
+def LAST : InstFlag<"printLast", 1>;
+
+def ADDRParam : ComplexPattern<i32, 2, "SelectADDRParam", [], []>;
+def ADDRDWord : ComplexPattern<i32, 1, "SelectADDRDWord", [], []>;
+def ADDRVTX_READ : ComplexPattern<i32, 2, "SelectADDRVTX_READ", [], []>;
+
+class R600ALU_Word0 {
+  field bits<32> Word0;
+
+  bits<11> src0;
+  bits<1>  src0_neg;
+  bits<1>  src0_rel;
+  bits<11> src1;
+  bits<1>  src1_rel;
+  bits<1>  src1_neg;
+  bits<3>  index_mode = 0;
+  bits<2>  pred_sel;
+  bits<1>  last;
+
+  bits<9>  src0_sel  = src0{8-0};
+  bits<2>  src0_chan = src0{10-9};
+  bits<9>  src1_sel  = src1{8-0};
+  bits<2>  src1_chan = src1{10-9};
+
+  let Word0{8-0}   = src0_sel;
+  let Word0{9}     = src0_rel;
+  let Word0{11-10} = src0_chan;
+  let Word0{12}    = src0_neg;
+  let Word0{21-13} = src1_sel;
+  let Word0{22}    = src1_rel;
+  let Word0{24-23} = src1_chan;
+  let Word0{25}    = src1_neg;
+  let Word0{28-26} = index_mode;
+  let Word0{30-29} = pred_sel;
+  let Word0{31}    = last;
+}
+
+class R600ALU_Word1 {
+  field bits<32> Word1;
+
+  bits<11> dst;
+  bits<3>  bank_swizzle = 0;
+  bits<1>  dst_rel;
+  bits<1>  clamp;
+
+  bits<7>  dst_sel  = dst{6-0};
+  bits<2>  dst_chan = dst{10-9};
+
+  let Word1{20-18} = bank_swizzle;
+  let Word1{27-21} = dst_sel;
+  let Word1{28}    = dst_rel;
+  let Word1{30-29} = dst_chan;
+  let Word1{31}    = clamp;
+}
+
+class R600ALU_Word1_OP2 <bits<11> alu_inst> : R600ALU_Word1{
+
+  bits<1>  src0_abs;
+  bits<1>  src1_abs;
+  bits<1>  update_exec_mask;
+  bits<1>  update_pred;
+  bits<1>  write;
+  bits<2>  omod;
+
+  let Word1{0}     = src0_abs;
+  let Word1{1}     = src1_abs;
+  let Word1{2}     = update_exec_mask;
+  let Word1{3}     = update_pred;
+  let Word1{4}     = write;
+  let Word1{6-5}   = omod;
+  let Word1{17-7}  = alu_inst;
+}
+
+class R600ALU_Word1_OP3 <bits<5> alu_inst> : R600ALU_Word1{
+
+  bits<11> src2;
+  bits<1>  src2_rel;
+  bits<1>  src2_neg;
+
+  bits<9>  src2_sel = src2{8-0};
+  bits<2>  src2_chan = src2{10-9};
+
+  let Word1{8-0}   = src2_sel;
+  let Word1{9}     = src2_rel;
+  let Word1{11-10} = src2_chan;
+  let Word1{12}    = src2_neg;
+  let Word1{17-13} = alu_inst;
+}
+
+/*
+XXX: R600 subtarget uses a slightly different encoding than the other
+subtargets.  We currently handle this in R600MCCodeEmitter, but we may
+want to use these instruction classes in the future.
+
+class R600ALU_Word1_OP2_r600 : R600ALU_Word1_OP2 {
+
+  bits<1>  fog_merge;
+  bits<10> alu_inst;
+
+  let Inst{37}    = fog_merge;
+  let Inst{39-38} = omod;
+  let Inst{49-40} = alu_inst;
+}
+
+class R600ALU_Word1_OP2_r700 : R600ALU_Word1_OP2 {
+
+  bits<11> alu_inst;
+
+  let Inst{38-37} = omod;
+  let Inst{49-39} = alu_inst;
+}
+*/
+
+def R600_Pred : PredicateOperand<i32, (ops R600_Predicate),
+                                     (ops PRED_SEL_OFF)>;
+
+
+let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
+
+// Class for instructions with only one source register.
+// If you add new ins to this instruction, make sure they are listed before
+// $literal, because the backend currently assumes that the last operand is
+// a literal.  Also be sure to update the enum R600Op1OperandIndex::ROI in
+// R600Defines.h, R600InstrInfo::buildDefaultInstruction(),
+// and R600InstrInfo::getOperandIdx().
+class R600_1OP <bits<11> inst, string opName, list<dag> pattern,
+                InstrItinClass itin = AnyALU> :
+    InstR600 <0,
+              (outs R600_Reg32:$dst),
+              (ins WRITE:$write, OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
+                   R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs,
+                   LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal),
+              !strconcat(opName,
+                   "$clamp $dst$write$dst_rel$omod, "
+                   "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
+                   "$literal $pred_sel$last"),
+              pattern,
+              itin>,
+    R600ALU_Word0,
+    R600ALU_Word1_OP2 <inst> {
+
+  let src1 = 0;
+  let src1_rel = 0;
+  let src1_neg = 0;
+  let src1_abs = 0;
+  let update_exec_mask = 0;
+  let update_pred = 0;
+  let HasNativeOperands = 1;
+  let Op1 = 1;
+  let DisableEncoding = "$literal";
+
+  let Inst{31-0}  = Word0;
+  let Inst{63-32} = Word1;
+}
+
+class R600_1OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
+                    InstrItinClass itin = AnyALU> :
+    R600_1OP <inst, opName,
+              [(set R600_Reg32:$dst, (node R600_Reg32:$src0))]
+>;
+
+// If you add our change the operands for R600_2OP instructions, you must
+// also update the R600Op2OperandIndex::ROI enum in R600Defines.h,
+// R600InstrInfo::buildDefaultInstruction(), and R600InstrInfo::getOperandIdx().
+class R600_2OP <bits<11> inst, string opName, list<dag> pattern,
+                InstrItinClass itin = AnyALU> :
+  InstR600 <inst,
+          (outs R600_Reg32:$dst),
+          (ins UEM:$update_exec_mask, UP:$update_pred, WRITE:$write,
+               OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
+               R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs,
+               R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, ABS:$src1_abs,
+               LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal),
+          !strconcat(opName,
+                "$clamp $update_exec_mask$update_pred$dst$write$dst_rel$omod, "
+                "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
+                "$src1_neg$src1_abs$src1$src1_abs$src1_rel, "
+                "$literal $pred_sel$last"),
+          pattern,
+          itin>,
+    R600ALU_Word0,
+    R600ALU_Word1_OP2 <inst> {
+
+  let HasNativeOperands = 1;
+  let Op2 = 1;
+  let DisableEncoding = "$literal";
+
+  let Inst{31-0}  = Word0;
+  let Inst{63-32} = Word1;
+}
+
+class R600_2OP_Helper <bits<11> inst, string opName, SDPatternOperator node,
+                       InstrItinClass itim = AnyALU> :
+    R600_2OP <inst, opName,
+              [(set R600_Reg32:$dst, (node R600_Reg32:$src0,
+                                           R600_Reg32:$src1))]
+>;
+
+// If you add our change the operands for R600_3OP instructions, you must
+// also update the R600Op3OperandIndex::ROI enum in R600Defines.h,
+// R600InstrInfo::buildDefaultInstruction(), and
+// R600InstrInfo::getOperandIdx().
+class R600_3OP <bits<5> inst, string opName, list<dag> pattern,
+                InstrItinClass itin = AnyALU> :
+  InstR600 <0,
+          (outs R600_Reg32:$dst),
+          (ins REL:$dst_rel, CLAMP:$clamp,
+               R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel,
+               R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel,
+               R600_Reg32:$src2, NEG:$src2_neg, REL:$src2_rel,
+               LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal),
+          !strconcat(opName, "$clamp $dst$dst_rel, "
+                             "$src0_neg$src0$src0_rel, "
+                             "$src1_neg$src1$src1_rel, "
+                             "$src2_neg$src2$src2_rel, "
+                             "$literal $pred_sel$last"),
+          pattern,
+          itin>,
+    R600ALU_Word0,
+    R600ALU_Word1_OP3<inst>{
+
+  let HasNativeOperands = 1;
+  let DisableEncoding = "$literal";
+  let Op3 = 1;
+
+  let Inst{31-0}  = Word0;
+  let Inst{63-32} = Word1;
+}
+
+class R600_REDUCTION <bits<11> inst, dag ins, string asm, list<dag> pattern,
+                      InstrItinClass itin = VecALU> :
+  InstR600 <inst,
+          (outs R600_Reg32:$dst),
+          ins,
+          asm,
+          pattern,
+          itin>;
+
+class R600_TEX <bits<11> inst, string opName, list<dag> pattern,
+                InstrItinClass itin = AnyALU> :
+  InstR600 <inst,
+          (outs R600_Reg128:$dst),
+          (ins R600_Reg128:$src0, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
+          !strconcat(opName, "$dst, $src0, $resourceId, $samplerId, $textureTarget"),
+          pattern,
+          itin>{
+    let Inst {10-0} = inst;
+  }
+
+} // End mayLoad = 1, mayStore = 0, hasSideEffects = 0
+
+def TEX_SHADOW : PatLeaf<
+  (imm),
+  [{uint32_t TType = (uint32_t)N->getZExtValue();
+    return (TType >= 6 && TType <= 8) || TType == 11 || TType == 12;
+  }]
+>;
+
+class EG_CF_RAT <bits <8> cf_inst, bits <6> rat_inst, bits<4> rat_id, dag outs,
+                 dag ins, string asm, list<dag> pattern> :
+    InstR600ISA <outs, ins, asm, pattern> {
+  bits<7>  RW_GPR;
+  bits<7>  INDEX_GPR;
+
+  bits<2>  RIM;
+  bits<2>  TYPE;
+  bits<1>  RW_REL;
+  bits<2>  ELEM_SIZE;
+
+  bits<12> ARRAY_SIZE;
+  bits<4>  COMP_MASK;
+  bits<4>  BURST_COUNT;
+  bits<1>  VPM;
+  bits<1>  eop;
+  bits<1>  MARK;
+  bits<1>  BARRIER;
+
+  // CF_ALLOC_EXPORT_WORD0_RAT
+  let Inst{3-0}   = rat_id;
+  let Inst{9-4}   = rat_inst;
+  let Inst{10}    = 0; // Reserved
+  let Inst{12-11} = RIM;
+  let Inst{14-13} = TYPE;
+  let Inst{21-15} = RW_GPR;
+  let Inst{22}    = RW_REL;
+  let Inst{29-23} = INDEX_GPR;
+  let Inst{31-30} = ELEM_SIZE;
+
+  // CF_ALLOC_EXPORT_WORD1_BUF
+  let Inst{43-32} = ARRAY_SIZE;
+  let Inst{47-44} = COMP_MASK;
+  let Inst{51-48} = BURST_COUNT;
+  let Inst{52}    = VPM;
+  let Inst{53}    = eop;
+  let Inst{61-54} = cf_inst;
+  let Inst{62}    = MARK;
+  let Inst{63}    = BARRIER;
+}
+
+class LoadParamFrag <PatFrag load_type> : PatFrag <
+  (ops node:$ptr), (load_type node:$ptr),
+  [{ return isParamLoad(dyn_cast<LoadSDNode>(N)); }]
+>;
+
+def load_param : LoadParamFrag<load>;
+def load_param_zexti8 : LoadParamFrag<zextloadi8>;
+def load_param_zexti16 : LoadParamFrag<zextloadi16>;
+
+def isR600 : Predicate<"Subtarget.device()"
+                            "->getGeneration() == AMDGPUDeviceInfo::HD4XXX">;
+def isR700 : Predicate<"Subtarget.device()"
+                            "->getGeneration() == AMDGPUDeviceInfo::HD4XXX &&"
+                            "Subtarget.device()->getDeviceFlag()"
+                            ">= OCL_DEVICE_RV710">;
+def isEG : Predicate<
+  "Subtarget.device()->getGeneration() >= AMDGPUDeviceInfo::HD5XXX && "
+  "Subtarget.device()->getGeneration() < AMDGPUDeviceInfo::HD7XXX && "
+  "Subtarget.device()->getDeviceFlag() != OCL_DEVICE_CAYMAN">;
+
+def isCayman : Predicate<"Subtarget.device()"
+                            "->getDeviceFlag() == OCL_DEVICE_CAYMAN">;
+def isEGorCayman : Predicate<"Subtarget.device()"
+                            "->getGeneration() == AMDGPUDeviceInfo::HD5XXX"
+                            "|| Subtarget.device()->getGeneration() =="
+                            "AMDGPUDeviceInfo::HD6XXX">;
+
+def isR600toCayman : Predicate<
+                     "Subtarget.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX">;
+
+//===----------------------------------------------------------------------===//
+// Interpolation Instructions
+//===----------------------------------------------------------------------===//
+
+def INTERP: SDNode<"AMDGPUISD::INTERP",
+  SDTypeProfile<1, 2, [SDTCisFP<0>, SDTCisInt<1>, SDTCisInt<2>]>
+  >;
+
+def INTERP_P0: SDNode<"AMDGPUISD::INTERP_P0",
+  SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisInt<1>]>
+  >;
+
+let usesCustomInserter = 1 in {
+def input_perspective :  AMDGPUShaderInst <
+  (outs R600_Reg128:$dst),
+  (ins i32imm:$src0, i32imm:$src1),
+  "input_perspective $src0 $src1 : dst",
+  [(set R600_Reg128:$dst, (INTERP (i32 imm:$src0), (i32 imm:$src1)))]>;
+}  // End usesCustomInserter = 1
+
+def input_constant :  AMDGPUShaderInst <
+  (outs R600_Reg128:$dst),
+  (ins i32imm:$src),
+  "input_perspective $src : dst",
+  [(set R600_Reg128:$dst, (INTERP_P0 (i32 imm:$src)))]>;
+
+
+
+def INTERP_XY : R600_2OP <0xD6, "INTERP_XY", []> {
+  let bank_swizzle = 5;
+}
+
+def INTERP_ZW : R600_2OP <0xD7, "INTERP_ZW", []> {
+  let bank_swizzle = 5;
+}
+
+def INTERP_LOAD_P0 : R600_1OP <0xE0, "INTERP_LOAD_P0", []>;
+
+//===----------------------------------------------------------------------===//
+// Export Instructions
+//===----------------------------------------------------------------------===//
+
+def ExportType : SDTypeProfile<0, 5, [SDTCisFP<0>, SDTCisInt<1>]>;
+
+def EXPORT: SDNode<"AMDGPUISD::EXPORT", ExportType,
+  [SDNPHasChain, SDNPSideEffect]>;
+
+class ExportWord0 {
+  field bits<32> Word0;
+
+  bits<13> arraybase;
+  bits<2> type;
+  bits<7> gpr;
+  bits<2> elem_size;
+
+  let Word0{12-0} = arraybase;
+  let Word0{14-13} = type;
+  let Word0{21-15} = gpr;
+  let Word0{22} = 0; // RW_REL
+  let Word0{29-23} = 0; // INDEX_GPR
+  let Word0{31-30} = elem_size;
+}
+
+class ExportSwzWord1 {
+  field bits<32> Word1;
+
+  bits<3> sw_x;
+  bits<3> sw_y;
+  bits<3> sw_z;
+  bits<3> sw_w;
+  bits<1> eop;
+  bits<8> inst;
+
+  let Word1{2-0} = sw_x;
+  let Word1{5-3} = sw_y;
+  let Word1{8-6} = sw_z;
+  let Word1{11-9} = sw_w;
+}
+
+class ExportBufWord1 {
+  field bits<32> Word1;
+
+  bits<12> arraySize;
+  bits<4> compMask;
+  bits<1> eop;
+  bits<8> inst;
+
+  let Word1{11-0} = arraySize;
+  let Word1{15-12} = compMask;
+}
+
+multiclass ExportPattern<Instruction ExportInst, bits<8> cf_inst> {
+  def : Pat<(int_R600_store_pixel_depth R600_Reg32:$reg),
+    (ExportInst
+        (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sel_x),
+        0, 61, 0, 7, 7, 7, cf_inst, 0)
+  >;
+
+  def : Pat<(int_R600_store_pixel_stencil R600_Reg32:$reg),
+    (ExportInst
+        (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), R600_Reg32:$reg, sel_x),
+        0, 61, 7, 0, 7, 7, cf_inst, 0)
+  >;
+
+  def : Pat<(int_R600_store_pixel_dummy),
+    (ExportInst
+        (v4f32 (IMPLICIT_DEF)), 0, 0, 7, 7, 7, 7, cf_inst, 0)
+  >;
+
+  def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 0),
+    (i32 imm:$type), (i32 imm:$arraybase), (i32 imm)),
+        (ExportInst R600_Reg128:$src, imm:$type, imm:$arraybase,
+        0, 1, 2, 3, cf_inst, 0)
+  >;
+}
+
+multiclass SteamOutputExportPattern<Instruction ExportInst,
+    bits<8> buf0inst, bits<8> buf1inst, bits<8> buf2inst, bits<8> buf3inst> {
+// Stream0
+  def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 1),
+      (i32 imm:$type), (i32 imm:$arraybase), (i32 imm:$mask)),
+      (ExportInst R600_Reg128:$src, imm:$type, imm:$arraybase,
+      4095, imm:$mask, buf0inst, 0)>;
+// Stream1
+  def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 2),
+      (i32 imm:$type), (i32 imm:$arraybase), (i32 imm:$mask)),
+      (ExportInst R600_Reg128:$src, imm:$type, imm:$arraybase,
+      4095, imm:$mask, buf1inst, 0)>;
+// Stream2
+  def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 3),
+      (i32 imm:$type), (i32 imm:$arraybase), (i32 imm:$mask)),
+      (ExportInst R600_Reg128:$src, imm:$type, imm:$arraybase,
+      4095, imm:$mask, buf2inst, 0)>;
+// Stream3
+  def : Pat<(EXPORT (v4f32 R600_Reg128:$src), (i32 4),
+      (i32 imm:$type), (i32 imm:$arraybase), (i32 imm:$mask)),
+      (ExportInst R600_Reg128:$src, imm:$type, imm:$arraybase,
+      4095, imm:$mask, buf3inst, 0)>;
+}
+
+let isTerminator = 1, usesCustomInserter = 1 in {
+
+class ExportSwzInst : InstR600ISA<(
+    outs),
+    (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
+    i32imm:$sw_x, i32imm:$sw_y, i32imm:$sw_z, i32imm:$sw_w, i32imm:$inst,
+    i32imm:$eop),
+    !strconcat("EXPORT", " $gpr"),
+    []>, ExportWord0, ExportSwzWord1 {
+  let elem_size = 3;
+  let Inst{31-0} = Word0;
+  let Inst{63-32} = Word1;
+}
+
+} // End isTerminator = 1, usesCustomInserter = 1
+
+class ExportBufInst : InstR600ISA<(
+    outs),
+    (ins R600_Reg128:$gpr, i32imm:$type, i32imm:$arraybase,
+    i32imm:$arraySize, i32imm:$compMask, i32imm:$inst, i32imm:$eop),
+    !strconcat("EXPORT", " $gpr"),
+    []>, ExportWord0, ExportBufWord1 {
+  let elem_size = 0;
+  let Inst{31-0} = Word0;
+  let Inst{63-32} = Word1;
+}
+
+let Predicates = [isR600toCayman] in { 
+
+//===----------------------------------------------------------------------===//
+// Common Instructions R600, R700, Evergreen, Cayman
+//===----------------------------------------------------------------------===//
+
+def ADD : R600_2OP_Helper <0x0, "ADD", fadd>;
+// Non-IEEE MUL: 0 * anything = 0
+def MUL : R600_2OP_Helper <0x1, "MUL NON-IEEE", int_AMDGPU_mul>;
+def MUL_IEEE : R600_2OP_Helper <0x2, "MUL_IEEE", fmul>;
+def MAX : R600_2OP_Helper <0x3, "MAX", AMDGPUfmax>;
+def MIN : R600_2OP_Helper <0x4, "MIN", AMDGPUfmin>;
+
+// For the SET* instructions there is a naming conflict in TargetSelectionDAG.td,
+// so some of the instruction names don't match the asm string.
+// XXX: Use the defs in TargetSelectionDAG.td instead of intrinsics.
+def SETE : R600_2OP <
+  0x08, "SETE",
+  [(set R600_Reg32:$dst,
+   (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
+             COND_EQ))]
+>;
+
+def SGT : R600_2OP <
+  0x09, "SETGT",
+  [(set R600_Reg32:$dst,
+   (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
+              COND_GT))]
+>;
+
+def SGE : R600_2OP <
+  0xA, "SETGE",
+  [(set R600_Reg32:$dst,
+   (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
+              COND_GE))]
+>;
+
+def SNE : R600_2OP <
+  0xB, "SETNE",
+  [(set R600_Reg32:$dst,
+   (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO,
+    COND_NE))]
+>;
+
+def FRACT : R600_1OP_Helper <0x10, "FRACT", AMDGPUfract>;
+def TRUNC : R600_1OP_Helper <0x11, "TRUNC", int_AMDGPU_trunc>;
+def CEIL : R600_1OP_Helper <0x12, "CEIL", fceil>;
+def RNDNE : R600_1OP_Helper <0x13, "RNDNE", frint>;
+def FLOOR : R600_1OP_Helper <0x14, "FLOOR", ffloor>;
+
+def MOV : R600_1OP <0x19, "MOV", []>;
+
+let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1 in {
+
+class MOV_IMM <ValueType vt, Operand immType> : AMDGPUInst <
+  (outs R600_Reg32:$dst),
+  (ins immType:$imm),
+  "",
+  []
+>;
+
+} // end let isPseudo = 1, isCodeGenOnly = 1, usesCustomInserter = 1
+
+def MOV_IMM_I32 : MOV_IMM<i32, i32imm>;
+def : Pat <
+  (imm:$val),
+  (MOV_IMM_I32 imm:$val)
+>;
+
+def MOV_IMM_F32 : MOV_IMM<f32, f32imm>;
+def : Pat <
+  (fpimm:$val),
+  (MOV_IMM_F32  fpimm:$val)
+>;
+
+def PRED_SETE : R600_2OP <0x20, "PRED_SETE", []>;
+def PRED_SETGT : R600_2OP <0x21, "PRED_SETGT", []>;
+def PRED_SETGE : R600_2OP <0x22, "PRED_SETGE", []>;
+def PRED_SETNE : R600_2OP <0x23, "PRED_SETNE", []>;
+
+let hasSideEffects = 1 in {
+
+def KILLGT : R600_2OP <0x2D, "KILLGT", []>;
+
+} // end hasSideEffects
+
+def AND_INT : R600_2OP_Helper <0x30, "AND_INT", and>;
+def OR_INT : R600_2OP_Helper <0x31, "OR_INT", or>;
+def XOR_INT : R600_2OP_Helper <0x32, "XOR_INT", xor>;
+def NOT_INT : R600_1OP_Helper <0x33, "NOT_INT", not>;
+def ADD_INT : R600_2OP_Helper <0x34, "ADD_INT", add>;
+def SUB_INT : R600_2OP_Helper <0x35, "SUB_INT", sub>;
+def MAX_INT : R600_2OP_Helper <0x36, "MAX_INT", AMDGPUsmax>;
+def MIN_INT : R600_2OP_Helper <0x37, "MIN_INT", AMDGPUsmin>;
+def MAX_UINT : R600_2OP_Helper <0x38, "MAX_UINT", AMDGPUsmax>;
+def MIN_UINT : R600_2OP_Helper <0x39, "MIN_UINT", AMDGPUumin>;
+
+def SETE_INT : R600_2OP <
+  0x3A, "SETE_INT",
+  [(set (i32 R600_Reg32:$dst),
+   (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETEQ))]
+>;
+
+def SETGT_INT : R600_2OP <
+  0x3B, "SGT_INT",
+  [(set (i32 R600_Reg32:$dst),
+   (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETGT))]
+>;
+
+def SETGE_INT : R600_2OP <
+  0x3C, "SETGE_INT",
+  [(set (i32 R600_Reg32:$dst),
+   (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETGE))]
+>;
+
+def SETNE_INT : R600_2OP <
+  0x3D, "SETNE_INT",
+  [(set (i32 R600_Reg32:$dst),
+   (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETNE))]
+>;
+
+def SETGT_UINT : R600_2OP <
+  0x3E, "SETGT_UINT",
+  [(set (i32 R600_Reg32:$dst),
+   (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUGT))]
+>;
+
+def SETGE_UINT : R600_2OP <
+  0x3F, "SETGE_UINT",
+  [(set (i32 R600_Reg32:$dst),
+    (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETUGE))]
+>;
+
+def PRED_SETE_INT : R600_2OP <0x42, "PRED_SETE_INT", []>;
+def PRED_SETGT_INT : R600_2OP <0x43, "PRED_SETGE_INT", []>;
+def PRED_SETGE_INT : R600_2OP <0x44, "PRED_SETGE_INT", []>;
+def PRED_SETNE_INT : R600_2OP <0x45, "PRED_SETNE_INT", []>;
+
+def CNDE_INT : R600_3OP <
+  0x1C, "CNDE_INT",
+  [(set (i32 R600_Reg32:$dst),
+   (selectcc (i32 R600_Reg32:$src0), 0,
+       (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
+       COND_EQ))]
+>;
+
+def CNDGE_INT : R600_3OP <
+  0x1E, "CNDGE_INT",
+  [(set (i32 R600_Reg32:$dst),
+   (selectcc (i32 R600_Reg32:$src0), 0,
+       (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
+       COND_GE))]
+>;
+
+def CNDGT_INT : R600_3OP <
+  0x1D, "CNDGT_INT",
+  [(set (i32 R600_Reg32:$dst),
+   (selectcc (i32 R600_Reg32:$src0), 0,
+       (i32 R600_Reg32:$src1), (i32 R600_Reg32:$src2),
+       COND_GT))]
+>;
+
+//===----------------------------------------------------------------------===//
+// Texture instructions
+//===----------------------------------------------------------------------===//
+
+def TEX_LD : R600_TEX <
+  0x03, "TEX_LD",
+  [(set R600_Reg128:$dst, (int_AMDGPU_txf R600_Reg128:$src0, imm:$src1, imm:$src2, imm:$src3, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+> {
+let AsmString = "TEX_LD $dst, $src0, $src1, $src2, $src3, $resourceId, $samplerId, $textureTarget";
+let InOperandList = (ins R600_Reg128:$src0, i32imm:$src1, i32imm:$src2, i32imm:$src3, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget);
+}
+
+def TEX_GET_TEXTURE_RESINFO : R600_TEX <
+  0x04, "TEX_GET_TEXTURE_RESINFO",
+  [(set R600_Reg128:$dst, (int_AMDGPU_txq R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+>;
+
+def TEX_GET_GRADIENTS_H : R600_TEX <
+  0x07, "TEX_GET_GRADIENTS_H",
+  [(set R600_Reg128:$dst, (int_AMDGPU_ddx R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+>;
+
+def TEX_GET_GRADIENTS_V : R600_TEX <
+  0x08, "TEX_GET_GRADIENTS_V",
+  [(set R600_Reg128:$dst, (int_AMDGPU_ddy R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+>;
+
+def TEX_SET_GRADIENTS_H : R600_TEX <
+  0x0B, "TEX_SET_GRADIENTS_H",
+  []
+>;
+
+def TEX_SET_GRADIENTS_V : R600_TEX <
+  0x0C, "TEX_SET_GRADIENTS_V",
+  []
+>;
+
+def TEX_SAMPLE : R600_TEX <
+  0x10, "TEX_SAMPLE",
+  [(set R600_Reg128:$dst, (int_AMDGPU_tex R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+>;
+
+def TEX_SAMPLE_C : R600_TEX <
+  0x18, "TEX_SAMPLE_C",
+  [(set R600_Reg128:$dst, (int_AMDGPU_tex R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
+>;
+
+def TEX_SAMPLE_L : R600_TEX <
+  0x11, "TEX_SAMPLE_L",
+  [(set R600_Reg128:$dst, (int_AMDGPU_txl R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+>;
+
+def TEX_SAMPLE_C_L : R600_TEX <
+  0x19, "TEX_SAMPLE_C_L",
+  [(set R600_Reg128:$dst, (int_AMDGPU_txl R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
+>;
+
+def TEX_SAMPLE_LB : R600_TEX <
+  0x12, "TEX_SAMPLE_LB",
+  [(set R600_Reg128:$dst, (int_AMDGPU_txb R600_Reg128:$src0,imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+>;
+
+def TEX_SAMPLE_C_LB : R600_TEX <
+  0x1A, "TEX_SAMPLE_C_LB",
+  [(set R600_Reg128:$dst, (int_AMDGPU_txb R600_Reg128:$src0, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
+>;
+
+def TEX_SAMPLE_G : R600_TEX <
+  0x14, "TEX_SAMPLE_G",
+  []
+>;
+
+def TEX_SAMPLE_C_G : R600_TEX <
+  0x1C, "TEX_SAMPLE_C_G",
+  []
+>;
+
+//===----------------------------------------------------------------------===//
+// Helper classes for common instructions
+//===----------------------------------------------------------------------===//
+
+class MUL_LIT_Common <bits<5> inst> : R600_3OP <
+  inst, "MUL_LIT",
+  []
+>;
+
+class MULADD_Common <bits<5> inst> : R600_3OP <
+  inst, "MULADD",
+  [(set (f32 R600_Reg32:$dst),
+   (IL_mad R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2))]
+>;
+
+class CNDE_Common <bits<5> inst> : R600_3OP <
+  inst, "CNDE",
+  [(set R600_Reg32:$dst,
+   (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
+       (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
+       COND_EQ))]
+>;
+
+class CNDGT_Common <bits<5> inst> : R600_3OP <
+  inst, "CNDGT",
+  [(set R600_Reg32:$dst,
+   (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
+       (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
+       COND_GT))]
+>;
+
+class CNDGE_Common <bits<5> inst> : R600_3OP <
+  inst, "CNDGE",
+  [(set R600_Reg32:$dst,
+   (selectcc (f32 R600_Reg32:$src0), FP_ZERO,
+       (f32 R600_Reg32:$src1), (f32 R600_Reg32:$src2),
+       COND_GE))]
+>;
+
+multiclass DOT4_Common <bits<11> inst> {
+
+  def _pseudo : R600_REDUCTION <inst,
+    (ins R600_Reg128:$src0, R600_Reg128:$src1),
+    "DOT4 $dst $src0, $src1",
+    [(set R600_Reg32:$dst, (int_AMDGPU_dp4 R600_Reg128:$src0, R600_Reg128:$src1))]
+  >;
+
+  def _real : R600_2OP <inst, "DOT4", []>;
+}
+
+let mayLoad = 0, mayStore = 0, hasSideEffects = 0 in {
+multiclass CUBE_Common <bits<11> inst> {
+
+  def _pseudo : InstR600 <
+    inst,
+    (outs R600_Reg128:$dst),
+    (ins R600_Reg128:$src),
+    "CUBE $dst $src",
+    [(set R600_Reg128:$dst, (int_AMDGPU_cube R600_Reg128:$src))],
+    VecALU
+  > {
+    let isPseudo = 1;
+  }
+
+  def _real : R600_2OP <inst, "CUBE", []>;
+}
+} // End mayLoad = 0, mayStore = 0, hasSideEffects = 0
+
+class EXP_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
+  inst, "EXP_IEEE", fexp2
+>;
+
+class FLT_TO_INT_Common <bits<11> inst> : R600_1OP_Helper <
+  inst, "FLT_TO_INT", fp_to_sint
+>;
+
+class INT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
+  inst, "INT_TO_FLT", sint_to_fp
+>;
+
+class FLT_TO_UINT_Common <bits<11> inst> : R600_1OP_Helper <
+  inst, "FLT_TO_UINT", fp_to_uint
+>;
+
+class UINT_TO_FLT_Common <bits<11> inst> : R600_1OP_Helper <
+  inst, "UINT_TO_FLT", uint_to_fp
+>;
+
+class LOG_CLAMPED_Common <bits<11> inst> : R600_1OP <
+  inst, "LOG_CLAMPED", []
+>;
+
+class LOG_IEEE_Common <bits<11> inst> : R600_1OP_Helper <
+  inst, "LOG_IEEE", flog2
+>;
+
+class LSHL_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHL", shl>;
+class LSHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "LSHR", srl>;
+class ASHR_Common <bits<11> inst> : R600_2OP_Helper <inst, "ASHR", sra>;
+class MULHI_INT_Common <bits<11> inst> : R600_2OP_Helper <
+  inst, "MULHI_INT", mulhs
+>;
+class MULHI_UINT_Common <bits<11> inst> : R600_2OP_Helper <
+  inst, "MULHI", mulhu
+>;
+class MULLO_INT_Common <bits<11> inst> : R600_2OP_Helper <
+  inst, "MULLO_INT", mul
+>;
+class MULLO_UINT_Common <bits<11> inst> : R600_2OP <inst, "MULLO_UINT", []>;
+
+class RECIP_CLAMPED_Common <bits<11> inst> : R600_1OP <
+  inst, "RECIP_CLAMPED", []
+>;
+
+class RECIP_IEEE_Common <bits<11> inst> : R600_1OP <
+  inst, "RECIP_IEEE", [(set R600_Reg32:$dst, (fdiv FP_ONE, R600_Reg32:$src0))]
+>;
+
+class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
+  inst, "RECIP_UINT", AMDGPUurecip
+>;
+
+class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
+  inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
+>;
+
+class RECIPSQRT_IEEE_Common <bits<11> inst> : R600_1OP <
+  inst, "RECIPSQRT_IEEE", []
+>;
+
+class SIN_Common <bits<11> inst> : R600_1OP <
+  inst, "SIN", []>{
+  let Trig = 1;
+}
+
+class COS_Common <bits<11> inst> : R600_1OP <
+  inst, "COS", []> {
+  let Trig = 1;
+}
+
+//===----------------------------------------------------------------------===//
+// Helper patterns for complex intrinsics
+//===----------------------------------------------------------------------===//
+
+multiclass DIV_Common <InstR600 recip_ieee> {
+def : Pat<
+  (int_AMDGPU_div R600_Reg32:$src0, R600_Reg32:$src1),
+  (MUL R600_Reg32:$src0, (recip_ieee R600_Reg32:$src1))
+>;
+
+def : Pat<
+  (fdiv R600_Reg32:$src0, R600_Reg32:$src1),
+  (MUL R600_Reg32:$src0, (recip_ieee R600_Reg32:$src1))
+>;
+}
+
+class TGSI_LIT_Z_Common <InstR600 mul_lit, InstR600 log_clamped, InstR600 exp_ieee> : Pat <
+  (int_TGSI_lit_z R600_Reg32:$src_x, R600_Reg32:$src_y, R600_Reg32:$src_w),
+  (exp_ieee (mul_lit (log_clamped (MAX R600_Reg32:$src_y, (f32 ZERO))), R600_Reg32:$src_w, R600_Reg32:$src_x))
+>;
+
+//===----------------------------------------------------------------------===//
+// R600 / R700 Instructions
+//===----------------------------------------------------------------------===//
+
+let Predicates = [isR600] in {
+
+  def MUL_LIT_r600 : MUL_LIT_Common<0x0C>;
+  def MULADD_r600 : MULADD_Common<0x10>;
+  def CNDE_r600 : CNDE_Common<0x18>;
+  def CNDGT_r600 : CNDGT_Common<0x19>;
+  def CNDGE_r600 : CNDGE_Common<0x1A>;
+  defm DOT4_r600 : DOT4_Common<0x50>;
+  defm CUBE_r600 : CUBE_Common<0x52>;
+  def EXP_IEEE_r600 : EXP_IEEE_Common<0x61>;
+  def LOG_CLAMPED_r600 : LOG_CLAMPED_Common<0x62>;
+  def LOG_IEEE_r600 : LOG_IEEE_Common<0x63>;
+  def RECIP_CLAMPED_r600 : RECIP_CLAMPED_Common<0x64>;
+  def RECIP_IEEE_r600 : RECIP_IEEE_Common<0x66>;
+  def RECIPSQRT_CLAMPED_r600 : RECIPSQRT_CLAMPED_Common<0x67>;
+  def RECIPSQRT_IEEE_r600 : RECIPSQRT_IEEE_Common<0x69>;
+  def FLT_TO_INT_r600 : FLT_TO_INT_Common<0x6b>;
+  def INT_TO_FLT_r600 : INT_TO_FLT_Common<0x6c>;
+  def FLT_TO_UINT_r600 : FLT_TO_UINT_Common<0x79>;
+  def UINT_TO_FLT_r600 : UINT_TO_FLT_Common<0x6d>;
+  def SIN_r600 : SIN_Common<0x6E>;
+  def COS_r600 : COS_Common<0x6F>;
+  def ASHR_r600 : ASHR_Common<0x70>;
+  def LSHR_r600 : LSHR_Common<0x71>;
+  def LSHL_r600 : LSHL_Common<0x72>;
+  def MULLO_INT_r600 : MULLO_INT_Common<0x73>;
+  def MULHI_INT_r600 : MULHI_INT_Common<0x74>;
+  def MULLO_UINT_r600 : MULLO_UINT_Common<0x75>;
+  def MULHI_UINT_r600 : MULHI_UINT_Common<0x76>;
+  def RECIP_UINT_r600 : RECIP_UINT_Common <0x78>;
+
+  defm DIV_r600 : DIV_Common<RECIP_IEEE_r600>;
+  def TGSI_LIT_Z_r600 : TGSI_LIT_Z_Common<MUL_LIT_r600, LOG_CLAMPED_r600, EXP_IEEE_r600>;
+
+  def : Pat<(fsqrt R600_Reg32:$src),
+    (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_r600 R600_Reg32:$src))>;
+
+  def R600_ExportSwz : ExportSwzInst {
+    let Word1{20-17} = 1; // BURST_COUNT
+    let Word1{21} = eop;
+    let Word1{22} = 1; // VALID_PIXEL_MODE
+    let Word1{30-23} = inst;
+    let Word1{31} = 1; // BARRIER
+  }
+  defm : ExportPattern<R600_ExportSwz, 39>;
+
+  def R600_ExportBuf : ExportBufInst {
+    let Word1{20-17} = 1; // BURST_COUNT
+    let Word1{21} = eop;
+    let Word1{22} = 1; // VALID_PIXEL_MODE
+    let Word1{30-23} = inst;
+    let Word1{31} = 1; // BARRIER
+  }
+  defm : SteamOutputExportPattern<R600_ExportBuf, 0x20, 0x21, 0x22, 0x23>;
+}
+
+// Helper pattern for normalizing inputs to triginomic instructions for R700+
+// cards.
+class COS_PAT <InstR600 trig> : Pat<
+  (fcos R600_Reg32:$src),
+  (trig (MUL (MOV_IMM_I32 CONST.TWO_PI_INV), R600_Reg32:$src))
+>;
+
+class SIN_PAT <InstR600 trig> : Pat<
+  (fsin R600_Reg32:$src),
+  (trig (MUL (MOV_IMM_I32 CONST.TWO_PI_INV), R600_Reg32:$src))
+>;
+
+//===----------------------------------------------------------------------===//
+// R700 Only instructions
+//===----------------------------------------------------------------------===//
+
+let Predicates = [isR700] in {
+  def SIN_r700 : SIN_Common<0x6E>;
+  def COS_r700 : COS_Common<0x6F>;
+
+  // R700 normalizes inputs to SIN/COS the same as EG
+  def : SIN_PAT <SIN_r700>;
+  def : COS_PAT <COS_r700>;
+}
+
+//===----------------------------------------------------------------------===//
+// Evergreen Only instructions
+//===----------------------------------------------------------------------===//
+
+let Predicates = [isEG] in {
+  
+def RECIP_IEEE_eg : RECIP_IEEE_Common<0x86>;
+defm DIV_eg : DIV_Common<RECIP_IEEE_eg>;
+
+def MULLO_INT_eg : MULLO_INT_Common<0x8F>;
+def MULHI_INT_eg : MULHI_INT_Common<0x90>;
+def MULLO_UINT_eg : MULLO_UINT_Common<0x91>;
+def MULHI_UINT_eg : MULHI_UINT_Common<0x92>;
+def RECIP_UINT_eg : RECIP_UINT_Common<0x94>;
+def RECIPSQRT_CLAMPED_eg : RECIPSQRT_CLAMPED_Common<0x87>;
+def EXP_IEEE_eg : EXP_IEEE_Common<0x81>;
+def LOG_IEEE_eg : LOG_IEEE_Common<0x83>;
+def RECIP_CLAMPED_eg : RECIP_CLAMPED_Common<0x84>;
+def RECIPSQRT_IEEE_eg : RECIPSQRT_IEEE_Common<0x89>;
+def SIN_eg : SIN_Common<0x8D>;
+def COS_eg : COS_Common<0x8E>;
+
+def : SIN_PAT <SIN_eg>;
+def : COS_PAT <COS_eg>;
+def : Pat<(fsqrt R600_Reg32:$src),
+  (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_eg R600_Reg32:$src))>;
+} // End Predicates = [isEG]
+
+//===----------------------------------------------------------------------===//
+// Evergreen / Cayman Instructions
+//===----------------------------------------------------------------------===//
+
+let Predicates = [isEGorCayman] in {
+
+  // BFE_UINT - bit_extract, an optimization for mask and shift
+  // Src0 = Input
+  // Src1 = Offset
+  // Src2 = Width
+  //
+  // bit_extract = (Input << (32 - Offset - Width)) >> (32 - Width)
+  //
+  // Example Usage:
+  // (Offset, Width)
+  //
+  // (0, 8)           = (Input << 24) >> 24  = (Input &  0xff)       >> 0
+  // (8, 8)           = (Input << 16) >> 24  = (Input &  0xffff)     >> 8
+  // (16,8)           = (Input <<  8) >> 24  = (Input &  0xffffff)   >> 16
+  // (24,8)           = (Input <<  0) >> 24  = (Input &  0xffffffff) >> 24
+  def BFE_UINT_eg : R600_3OP <0x4, "BFE_UINT",
+    [(set R600_Reg32:$dst, (int_AMDIL_bit_extract_u32 R600_Reg32:$src0,
+                                                      R600_Reg32:$src1,
+                                                      R600_Reg32:$src2))],
+    VecALU
+  >;
+
+  def BIT_ALIGN_INT_eg : R600_3OP <0xC, "BIT_ALIGN_INT",
+    [(set R600_Reg32:$dst, (AMDGPUbitalign R600_Reg32:$src0, R600_Reg32:$src1,
+                                          R600_Reg32:$src2))],
+    VecALU
+  >;
+
+  def MULADD_eg : MULADD_Common<0x14>;
+  def ASHR_eg : ASHR_Common<0x15>;
+  def LSHR_eg : LSHR_Common<0x16>;
+  def LSHL_eg : LSHL_Common<0x17>;
+  def CNDE_eg : CNDE_Common<0x19>;
+  def CNDGT_eg : CNDGT_Common<0x1A>;
+  def CNDGE_eg : CNDGE_Common<0x1B>;
+  def MUL_LIT_eg : MUL_LIT_Common<0x1F>;
+  def LOG_CLAMPED_eg : LOG_CLAMPED_Common<0x82>;
+  defm DOT4_eg : DOT4_Common<0xBE>;
+  defm CUBE_eg : CUBE_Common<0xC0>;
+
+  def TGSI_LIT_Z_eg : TGSI_LIT_Z_Common<MUL_LIT_eg, LOG_CLAMPED_eg, EXP_IEEE_eg>;
+
+  def FLT_TO_INT_eg : FLT_TO_INT_Common<0x50> {
+    let Pattern = [];
+  }
+
+  def INT_TO_FLT_eg : INT_TO_FLT_Common<0x9B>;
+
+  def FLT_TO_UINT_eg : FLT_TO_UINT_Common<0x9A> {
+    let Pattern = [];
+  }
+
+  def UINT_TO_FLT_eg : UINT_TO_FLT_Common<0x9C>;
+
+  // TRUNC is used for the FLT_TO_INT instructions to work around a
+  // perceived problem where the rounding modes are applied differently
+  // depending on the instruction and the slot they are in.
+  // See:
+  // https://bugs.freedesktop.org/show_bug.cgi?id=50232
+  // Mesa commit: a1a0974401c467cb86ef818f22df67c21774a38c
+  //
+  // XXX: Lowering SELECT_CC will sometimes generate fp_to_[su]int nodes,
+  // which do not need to be truncated since the fp values are 0.0f or 1.0f.
+  // We should look into handling these cases separately.
+  def : Pat<(fp_to_sint R600_Reg32:$src0),
+    (FLT_TO_INT_eg (TRUNC R600_Reg32:$src0))>;
+
+  def : Pat<(fp_to_uint R600_Reg32:$src0),
+    (FLT_TO_UINT_eg (TRUNC R600_Reg32:$src0))>;
+
+  def EG_ExportSwz : ExportSwzInst {
+    let Word1{19-16} = 1; // BURST_COUNT
+    let Word1{20} = 1; // VALID_PIXEL_MODE
+    let Word1{21} = eop;
+    let Word1{29-22} = inst;
+    let Word1{30} = 0; // MARK
+    let Word1{31} = 1; // BARRIER
+  }
+  defm : ExportPattern<EG_ExportSwz, 83>;
+
+  def EG_ExportBuf : ExportBufInst {
+    let Word1{19-16} = 1; // BURST_COUNT
+    let Word1{20} = 1; // VALID_PIXEL_MODE
+    let Word1{21} = eop;
+    let Word1{29-22} = inst;
+    let Word1{30} = 0; // MARK
+    let Word1{31} = 1; // BARRIER
+  }
+  defm : SteamOutputExportPattern<EG_ExportBuf, 0x40, 0x41, 0x42, 0x43>;
+
+//===----------------------------------------------------------------------===//
+// Memory read/write instructions
+//===----------------------------------------------------------------------===//
+let usesCustomInserter = 1 in {
+
+class RAT_WRITE_CACHELESS_eg <dag ins, bits<4> comp_mask, string name,
+                              list<dag> pattern>
+    : EG_CF_RAT <0x57, 0x2, 0, (outs), ins,
+                 !strconcat(name, " $rw_gpr, $index_gpr, $eop"), pattern> {
+  let RIM         = 0;
+  // XXX: Have a separate instruction for non-indexed writes.
+  let TYPE        = 1;
+  let RW_REL      = 0;
+  let ELEM_SIZE   = 0;
+
+  let ARRAY_SIZE  = 0;
+  let COMP_MASK   = comp_mask;
+  let BURST_COUNT = 0;
+  let VPM         = 0;
+  let MARK        = 0;
+  let BARRIER     = 1;
+}
+
+} // End usesCustomInserter = 1
+
+// 32-bit store
+def RAT_WRITE_CACHELESS_32_eg : RAT_WRITE_CACHELESS_eg <
+  (ins R600_TReg32_X:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
+  0x1, "RAT_WRITE_CACHELESS_32_eg",
+  [(global_store (i32 R600_TReg32_X:$rw_gpr), R600_TReg32_X:$index_gpr)]
+>;
+
+//128-bit store
+def RAT_WRITE_CACHELESS_128_eg : RAT_WRITE_CACHELESS_eg <
+  (ins R600_Reg128:$rw_gpr, R600_TReg32_X:$index_gpr, InstFlag:$eop),
+  0xf, "RAT_WRITE_CACHELESS_128",
+  [(global_store (v4i32 R600_Reg128:$rw_gpr), R600_TReg32_X:$index_gpr)]
+>;
+
+class VTX_READ_eg <string name, bits<8> buffer_id, dag outs, list<dag> pattern>
+    : InstR600ISA <outs, (ins MEMxi:$ptr), name#" $dst, $ptr", pattern> {
+
+  // Operands
+  bits<7> DST_GPR;
+  bits<7> SRC_GPR;
+
+  // Static fields
+  bits<5> VC_INST = 0;
+  bits<2> FETCH_TYPE = 2;
+  bits<1> FETCH_WHOLE_QUAD = 0;
+  bits<8> BUFFER_ID = buffer_id;
+  bits<1> SRC_REL = 0;
+  // XXX: We can infer this field based on the SRC_GPR.  This would allow us
+  // to store vertex addresses in any channel, not just X.
+  bits<2> SRC_SEL_X = 0;
+  bits<6> MEGA_FETCH_COUNT;
+  bits<1> DST_REL = 0;
+  bits<3> DST_SEL_X;
+  bits<3> DST_SEL_Y;
+  bits<3> DST_SEL_Z;
+  bits<3> DST_SEL_W;
+  // The docs say that if this bit is set, then DATA_FORMAT, NUM_FORMAT_ALL,
+  // FORMAT_COMP_ALL, SRF_MODE_ALL, and ENDIAN_SWAP fields will be ignored,
+  // however, based on my testing if USE_CONST_FIELDS is set, then all
+  // these fields need to be set to 0.
+  bits<1> USE_CONST_FIELDS = 0;
+  bits<6> DATA_FORMAT;
+  bits<2> NUM_FORMAT_ALL = 1;
+  bits<1> FORMAT_COMP_ALL = 0;
+  bits<1> SRF_MODE_ALL = 0;
+
+  // LLVM can only encode 64-bit instructions, so these fields are manually
+  // encoded in R600CodeEmitter
+  //
+  // bits<16> OFFSET;
+  // bits<2>  ENDIAN_SWAP = 0;
+  // bits<1>  CONST_BUF_NO_STRIDE = 0;
+  // bits<1>  MEGA_FETCH = 0;
+  // bits<1>  ALT_CONST = 0;
+  // bits<2>  BUFFER_INDEX_MODE = 0;
+
+  // VTX_WORD0
+  let Inst{4-0}   = VC_INST;
+  let Inst{6-5}   = FETCH_TYPE;
+  let Inst{7}     = FETCH_WHOLE_QUAD;
+  let Inst{15-8}  = BUFFER_ID;
+  let Inst{22-16} = SRC_GPR;
+  let Inst{23}    = SRC_REL;
+  let Inst{25-24} = SRC_SEL_X;
+  let Inst{31-26} = MEGA_FETCH_COUNT;
+
+  // VTX_WORD1_GPR
+  let Inst{38-32} = DST_GPR;
+  let Inst{39}    = DST_REL;
+  let Inst{40}    = 0; // Reserved
+  let Inst{43-41} = DST_SEL_X;
+  let Inst{46-44} = DST_SEL_Y;
+  let Inst{49-47} = DST_SEL_Z;
+  let Inst{52-50} = DST_SEL_W;
+  let Inst{53}    = USE_CONST_FIELDS;
+  let Inst{59-54} = DATA_FORMAT;
+  let Inst{61-60} = NUM_FORMAT_ALL;
+  let Inst{62}    = FORMAT_COMP_ALL;
+  let Inst{63}    = SRF_MODE_ALL;
+
+  // VTX_WORD2 (LLVM can only encode 64-bit instructions, so WORD2 encoding
+  // is done in R600CodeEmitter
+  //
+  // Inst{79-64} = OFFSET;
+  // Inst{81-80} = ENDIAN_SWAP;
+  // Inst{82}    = CONST_BUF_NO_STRIDE;
+  // Inst{83}    = MEGA_FETCH;
+  // Inst{84}    = ALT_CONST;
+  // Inst{86-85} = BUFFER_INDEX_MODE;
+  // Inst{95-86} = 0; Reserved
+
+  // VTX_WORD3 (Padding)
+  //
+  // Inst{127-96} = 0;
+}
+
+class VTX_READ_8_eg <bits<8> buffer_id, list<dag> pattern>
+    : VTX_READ_eg <"VTX_READ_8", buffer_id, (outs R600_TReg32_X:$dst),
+                   pattern> {
+
+  let MEGA_FETCH_COUNT = 1;
+  let DST_SEL_X = 0;
+  let DST_SEL_Y = 7;   // Masked
+  let DST_SEL_Z = 7;   // Masked
+  let DST_SEL_W = 7;   // Masked
+  let DATA_FORMAT = 1; // FMT_8
+}
+
+class VTX_READ_16_eg <bits<8> buffer_id, list<dag> pattern>
+    : VTX_READ_eg <"VTX_READ_16", buffer_id, (outs R600_TReg32_X:$dst),
+                    pattern> {
+  let MEGA_FETCH_COUNT = 2;
+  let DST_SEL_X = 0;
+  let DST_SEL_Y = 7;   // Masked
+  let DST_SEL_Z = 7;   // Masked
+  let DST_SEL_W = 7;   // Masked
+  let DATA_FORMAT = 5; // FMT_16
+
+}
+
+class VTX_READ_32_eg <bits<8> buffer_id, list<dag> pattern>
+    : VTX_READ_eg <"VTX_READ_32", buffer_id, (outs R600_TReg32_X:$dst),
+                   pattern> {
+
+  let MEGA_FETCH_COUNT = 4;
+  let DST_SEL_X        = 0;
+  let DST_SEL_Y        = 7;   // Masked
+  let DST_SEL_Z        = 7;   // Masked
+  let DST_SEL_W        = 7;   // Masked
+  let DATA_FORMAT      = 0xD; // COLOR_32
+
+  // This is not really necessary, but there were some GPU hangs that appeared
+  // to be caused by ALU instructions in the next instruction group that wrote
+  // to the $ptr registers of the VTX_READ.  
+  // e.g.
+  // %T3_X<def> = VTX_READ_PARAM_32_eg %T2_X<kill>, 24
+  // %T2_X<def> = MOV %ZERO
+  //Adding this constraint prevents this from happening.
+  let Constraints = "$ptr.ptr = $dst";
+}
+
+class VTX_READ_128_eg <bits<8> buffer_id, list<dag> pattern>
+    : VTX_READ_eg <"VTX_READ_128", buffer_id, (outs R600_Reg128:$dst),
+                   pattern> {
+
+  let MEGA_FETCH_COUNT = 16;
+  let DST_SEL_X        =  0;
+  let DST_SEL_Y        =  1;
+  let DST_SEL_Z        =  2;
+  let DST_SEL_W        =  3;
+  let DATA_FORMAT      =  0x22; // COLOR_32_32_32_32
+
+  // XXX: Need to force VTX_READ_128 instructions to write to the same register
+  // that holds its buffer address to avoid potential hangs.  We can't use
+  // the same constraint as VTX_READ_32_eg, because the $ptr.ptr and $dst
+  // registers are different sizes.
+}
+
+//===----------------------------------------------------------------------===//
+// VTX Read from parameter memory space
+//===----------------------------------------------------------------------===//
+
+def VTX_READ_PARAM_8_eg : VTX_READ_8_eg <0,
+  [(set (i32 R600_TReg32_X:$dst), (load_param_zexti8 ADDRVTX_READ:$ptr))]
+>;
+
+def VTX_READ_PARAM_16_eg : VTX_READ_16_eg <0,
+  [(set (i32 R600_TReg32_X:$dst), (load_param_zexti16 ADDRVTX_READ:$ptr))]
+>;
+
+def VTX_READ_PARAM_32_eg : VTX_READ_32_eg <0,
+  [(set (i32 R600_TReg32_X:$dst), (load_param ADDRVTX_READ:$ptr))]
+>;
+
+//===----------------------------------------------------------------------===//
+// VTX Read from global memory space
+//===----------------------------------------------------------------------===//
+
+// 8-bit reads
+def VTX_READ_GLOBAL_8_eg : VTX_READ_8_eg <1,
+  [(set (i32 R600_TReg32_X:$dst), (zextloadi8_global ADDRVTX_READ:$ptr))]
+>;
+
+// 32-bit reads
+def VTX_READ_GLOBAL_32_eg : VTX_READ_32_eg <1,
+  [(set (i32 R600_TReg32_X:$dst), (global_load ADDRVTX_READ:$ptr))]
+>;
+
+// 128-bit reads
+def VTX_READ_GLOBAL_128_eg : VTX_READ_128_eg <1,
+  [(set (v4i32 R600_Reg128:$dst), (global_load ADDRVTX_READ:$ptr))]
+>;
+
+//===----------------------------------------------------------------------===//
+// Constant Loads
+// XXX: We are currently storing all constants in the global address space.
+//===----------------------------------------------------------------------===//
+
+def CONSTANT_LOAD_eg : VTX_READ_32_eg <1,
+  [(set (i32 R600_TReg32_X:$dst), (constant_load ADDRVTX_READ:$ptr))]
+>;
+
+}
+
+let Predicates = [isCayman] in {
+
+let isVector = 1 in { 
+
+def RECIP_IEEE_cm : RECIP_IEEE_Common<0x86>;
+
+def MULLO_INT_cm : MULLO_INT_Common<0x8F>;
+def MULHI_INT_cm : MULHI_INT_Common<0x90>;
+def MULLO_UINT_cm : MULLO_UINT_Common<0x91>;
+def MULHI_UINT_cm : MULHI_UINT_Common<0x92>;
+def RECIPSQRT_CLAMPED_cm : RECIPSQRT_CLAMPED_Common<0x87>;
+def EXP_IEEE_cm : EXP_IEEE_Common<0x81>;
+def LOG_IEEE_ : LOG_IEEE_Common<0x83>;
+def RECIP_CLAMPED_cm : RECIP_CLAMPED_Common<0x84>;
+def RECIPSQRT_IEEE_cm : RECIPSQRT_IEEE_Common<0x89>;
+def SIN_cm : SIN_Common<0x8D>;
+def COS_cm : COS_Common<0x8E>;
+} // End isVector = 1
+
+def : SIN_PAT <SIN_cm>;
+def : COS_PAT <COS_cm>;
+
+defm DIV_cm : DIV_Common<RECIP_IEEE_cm>;
+
+// RECIP_UINT emulation for Cayman
+def : Pat <
+  (AMDGPUurecip R600_Reg32:$src0),
+  (FLT_TO_UINT_eg (MUL_IEEE (RECIP_IEEE_cm (UINT_TO_FLT_eg R600_Reg32:$src0)),
+                            (MOV_IMM_I32 0x4f800000)))
+>;
+
+
+def : Pat<(fsqrt R600_Reg32:$src),
+  (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm R600_Reg32:$src))>;
+
+} // End isCayman
+
+//===----------------------------------------------------------------------===//
+// Branch Instructions
+//===----------------------------------------------------------------------===//
+
+
+def IF_PREDICATE_SET  : ILFormat<(outs), (ins GPRI32:$src),
+  "IF_PREDICATE_SET $src", []>;
+
+def PREDICATED_BREAK : ILFormat<(outs), (ins GPRI32:$src),
+  "PREDICATED_BREAK $src", []>;
+
+//===----------------------------------------------------------------------===//
+// Pseudo instructions
+//===----------------------------------------------------------------------===//
+
+let isPseudo = 1 in {
+
+def PRED_X : InstR600 <
+  0, (outs R600_Predicate_Bit:$dst),
+  (ins R600_Reg32:$src0, i32imm:$src1, i32imm:$flags),
+  "", [], NullALU> {
+  let FlagOperandIdx = 3;
+}
+
+let isTerminator = 1, isBranch = 1, isBarrier = 1 in {
+
+def JUMP : InstR600 <0x10,
+          (outs),
+          (ins brtarget:$target, R600_Pred:$p),
+          "JUMP $target ($p)",
+          [], AnyALU
+  >;
+
+}  // End isTerminator = 1, isBranch = 1, isBarrier = 1
+
+let usesCustomInserter = 1 in {
+
+let mayLoad = 0, mayStore = 0, hasSideEffects = 1 in {
+
+def MASK_WRITE : AMDGPUShaderInst <
+    (outs),
+    (ins R600_Reg32:$src),
+    "MASK_WRITE $src",
+    []
+>;
+
+} // End mayLoad = 0, mayStore = 0, hasSideEffects = 1
+
+def R600_LOAD_CONST : AMDGPUShaderInst <
+  (outs R600_Reg32:$dst),
+  (ins i32imm:$src0),
+  "R600_LOAD_CONST $dst, $src0",
+  [(set R600_Reg32:$dst, (int_AMDGPU_load_const imm:$src0))]
+>;
+
+def RESERVE_REG : AMDGPUShaderInst <
+  (outs),
+  (ins i32imm:$src),
+  "RESERVE_REG $src",
+  [(int_AMDGPU_reserve_reg imm:$src)]
+>;
+
+def TXD: AMDGPUShaderInst <
+  (outs R600_Reg128:$dst),
+  (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
+  "TXD $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
+  [(set R600_Reg128:$dst, (int_AMDGPU_txd R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, imm:$resourceId, imm:$samplerId, imm:$textureTarget))]
+>;
+
+def TXD_SHADOW: AMDGPUShaderInst <
+  (outs R600_Reg128:$dst),
+  (ins R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, i32imm:$resourceId, i32imm:$samplerId, i32imm:$textureTarget),
+  "TXD_SHADOW $dst, $src0, $src1, $src2, $resourceId, $samplerId, $textureTarget",
+  [(set R600_Reg128:$dst, (int_AMDGPU_txd R600_Reg128:$src0, R600_Reg128:$src1, R600_Reg128:$src2, imm:$resourceId, imm:$samplerId, TEX_SHADOW:$textureTarget))]
+>;
+
+} // End isPseudo = 1
+} // End usesCustomInserter = 1
+
+def CLAMP_R600 :  CLAMP <R600_Reg32>;
+def FABS_R600 : FABS<R600_Reg32>;
+def FNEG_R600 : FNEG<R600_Reg32>;
+
+//===---------------------------------------------------------------------===//
+// Return instruction
+//===---------------------------------------------------------------------===//
+let isTerminator = 1, isReturn = 1, isBarrier = 1, hasCtrlDep = 1 in {
+  def RETURN          : ILFormat<(outs), (ins variable_ops),
+      "RETURN", [(IL_retflag)]>;
+}
+
+//===----------------------------------------------------------------------===//
+// ISel Patterns
+//===----------------------------------------------------------------------===//
+
+//CNDGE_INT extra pattern
+def : Pat <
+  (selectcc (i32 R600_Reg32:$src0), -1, (i32 R600_Reg32:$src1),
+                                        (i32 R600_Reg32:$src2), COND_GT),
+  (CNDGE_INT R600_Reg32:$src0, R600_Reg32:$src1, R600_Reg32:$src2)
+>;
+
+// KIL Patterns
+def KILP : Pat <
+  (int_AMDGPU_kilp),
+  (MASK_WRITE (KILLGT (f32 ONE), (f32 ZERO)))
+>;
+
+def KIL : Pat <
+  (int_AMDGPU_kill R600_Reg32:$src0),
+  (MASK_WRITE (KILLGT (f32 ZERO), (f32 R600_Reg32:$src0)))
+>;
+
+// SGT Reverse args
+def : Pat <
+  (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, COND_LT),
+  (SGT R600_Reg32:$src1, R600_Reg32:$src0)
+>;
+
+// SGE Reverse args
+def : Pat <
+  (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, COND_LE),
+  (SGE R600_Reg32:$src1, R600_Reg32:$src0) 
+>;
+
+// SETGT_INT reverse args
+def : Pat <
+  (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETLT),
+  (SETGT_INT R600_Reg32:$src1, R600_Reg32:$src0)
+>;
+
+// SETGE_INT reverse args
+def : Pat <
+  (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETLE),
+  (SETGE_INT R600_Reg32:$src1, R600_Reg32:$src0)
+>;
+
+// SETGT_UINT reverse args
+def : Pat <
+  (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETULT),
+  (SETGT_UINT R600_Reg32:$src1, R600_Reg32:$src0)
+>;
+
+// SETGE_UINT reverse args
+def : Pat <
+  (selectcc (i32 R600_Reg32:$src0), R600_Reg32:$src1, -1, 0, SETULE),
+  (SETGE_UINT R600_Reg32:$src1, R600_Reg32:$src0)
+>;
+
+// The next two patterns are special cases for handling 'true if ordered' and
+// 'true if unordered' conditionals.  The assumption here is that the behavior of
+// SETE and SNE conforms to the Direct3D 10 rules for floating point values
+// described here:
+// http://msdn.microsoft.com/en-us/library/windows/desktop/cc308050.aspx#alpha_32_bit
+// We assume that  SETE returns false when one of the operands is NAN and
+// SNE returns true when on of the operands is NAN
+
+//SETE - 'true if ordered'
+def : Pat <
+  (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, SETO),
+  (SETE R600_Reg32:$src0, R600_Reg32:$src1)
+>;
+
+//SNE - 'true if unordered'
+def : Pat <
+  (selectcc (f32 R600_Reg32:$src0), R600_Reg32:$src1, FP_ONE, FP_ZERO, SETUO),
+  (SNE R600_Reg32:$src0, R600_Reg32:$src1)
+>;
+
+def : Extract_Element <f32, v4f32, R600_Reg128, 0, sel_x>;
+def : Extract_Element <f32, v4f32, R600_Reg128, 1, sel_y>;
+def : Extract_Element <f32, v4f32, R600_Reg128, 2, sel_z>;
+def : Extract_Element <f32, v4f32, R600_Reg128, 3, sel_w>;
+
+def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 0, sel_x>;
+def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 1, sel_y>;
+def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 2, sel_z>;
+def : Insert_Element <f32, v4f32, R600_Reg32, R600_Reg128, 3, sel_w>;
+
+def : Extract_Element <i32, v4i32, R600_Reg128, 0, sel_x>;
+def : Extract_Element <i32, v4i32, R600_Reg128, 1, sel_y>;
+def : Extract_Element <i32, v4i32, R600_Reg128, 2, sel_z>;
+def : Extract_Element <i32, v4i32, R600_Reg128, 3, sel_w>;
+
+def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 0, sel_x>;
+def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 1, sel_y>;
+def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 2, sel_z>;
+def : Insert_Element <i32, v4i32, R600_Reg32, R600_Reg128, 3, sel_w>;
+
+def : Vector_Build <v4f32, R600_Reg128, f32, R600_Reg32>;
+def : Vector_Build <v4i32, R600_Reg128, i32, R600_Reg32>;
+
+// bitconvert patterns
+
+def : BitConvert <i32, f32, R600_Reg32>;
+def : BitConvert <f32, i32, R600_Reg32>;
+def : BitConvert <v4f32, v4i32, R600_Reg128>;
+def : BitConvert <v4i32, v4f32, R600_Reg128>;
+
+// DWORDADDR pattern
+def : DwordAddrPat  <i32, R600_Reg32>;
+
+} // End isR600toCayman Predicate
diff --git a/lib/Target/R600/R600Intrinsics.td b/lib/Target/R600/R600Intrinsics.td
new file mode 100644
index 0000000000..3825bc4d3b
--- /dev/null
+++ b/lib/Target/R600/R600Intrinsics.td
@@ -0,0 +1,32 @@
+//===-- R600Intrinsics.td - R600 Instrinsic defs -------*- tablegen -*-----===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// R600 Intrinsic Definitions
+//
+//===----------------------------------------------------------------------===//
+
+let TargetPrefix = "R600", isTarget = 1 in {
+  def int_R600_load_input : Intrinsic<[llvm_float_ty], [llvm_i32_ty], [IntrNoMem]>;
+  def int_R600_load_input_perspective :
+    Intrinsic<[llvm_float_ty], [llvm_i32_ty], [IntrReadMem]>;
+  def int_R600_load_input_constant :
+    Intrinsic<[llvm_float_ty], [llvm_i32_ty], [IntrReadMem]>;
+  def int_R600_load_input_linear :
+    Intrinsic<[llvm_float_ty], [llvm_i32_ty], [IntrReadMem]>;
+  def int_R600_store_stream_output :
+    Intrinsic<[], [llvm_float_ty, llvm_i32_ty, llvm_i32_ty], []>;
+  def int_R600_store_pixel_color :
+      Intrinsic<[], [llvm_float_ty, llvm_i32_ty], []>;
+  def int_R600_store_pixel_depth :
+      Intrinsic<[], [llvm_float_ty], []>;
+  def int_R600_store_pixel_stencil :
+      Intrinsic<[], [llvm_float_ty], []>;
+  def int_R600_store_pixel_dummy :
+      Intrinsic<[], [], []>;
+}
diff --git a/lib/Target/R600/R600MachineFunctionInfo.cpp b/lib/Target/R600/R600MachineFunctionInfo.cpp
new file mode 100644
index 0000000000..4eb5efa19f
--- /dev/null
+++ b/lib/Target/R600/R600MachineFunctionInfo.cpp
@@ -0,0 +1,34 @@
+//===-- R600MachineFunctionInfo.cpp - R600 Machine Function Info-*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//===----------------------------------------------------------------------===//
+
+#include "R600MachineFunctionInfo.h"
+
+using namespace llvm;
+
+R600MachineFunctionInfo::R600MachineFunctionInfo(const MachineFunction &MF)
+  : MachineFunctionInfo(),
+    HasLinearInterpolation(false),
+    HasPerspectiveInterpolation(false) {
+    memset(Outputs, 0, sizeof(Outputs));
+    memset(StreamOutputs, 0, sizeof(StreamOutputs));
+  }
+
+unsigned R600MachineFunctionInfo::GetIJPerspectiveIndex() const {
+  assert(HasPerspectiveInterpolation);
+  return 0;
+}
+
+unsigned R600MachineFunctionInfo::GetIJLinearIndex() const {
+  assert(HasLinearInterpolation);
+  if (HasPerspectiveInterpolation)
+    return 1;
+  else
+    return 0;
+}
diff --git a/lib/Target/R600/R600MachineFunctionInfo.h b/lib/Target/R600/R600MachineFunctionInfo.h
new file mode 100644
index 0000000000..e97fb5be62
--- /dev/null
+++ b/lib/Target/R600/R600MachineFunctionInfo.h
@@ -0,0 +1,39 @@
+//===-- R600MachineFunctionInfo.h - R600 Machine Function Info ----*- C++ -*-=//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+//===----------------------------------------------------------------------===//
+
+#ifndef R600MACHINEFUNCTIONINFO_H
+#define R600MACHINEFUNCTIONINFO_H
+
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/SelectionDAG.h"
+#include <vector>
+
+namespace llvm {
+
+class R600MachineFunctionInfo : public MachineFunctionInfo {
+
+public:
+  R600MachineFunctionInfo(const MachineFunction &MF);
+  std::vector<unsigned> ReservedRegs;
+  SDNode *Outputs[16];
+  SDNode *StreamOutputs[64][4];
+  bool HasLinearInterpolation;
+  bool HasPerspectiveInterpolation;
+
+  unsigned GetIJLinearIndex() const;
+  unsigned GetIJPerspectiveIndex() const;
+
+};
+
+} // End llvm namespace
+
+#endif //R600MACHINEFUNCTIONINFO_H
diff --git a/lib/Target/R600/R600RegisterInfo.cpp b/lib/Target/R600/R600RegisterInfo.cpp
new file mode 100644
index 0000000000..a39f83dbac
--- /dev/null
+++ b/lib/Target/R600/R600RegisterInfo.cpp
@@ -0,0 +1,89 @@
+//===-- R600RegisterInfo.cpp - R600 Register Information ------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief R600 implementation of the TargetRegisterInfo class.
+//
+//===----------------------------------------------------------------------===//
+
+#include "R600RegisterInfo.h"
+#include "AMDGPUTargetMachine.h"
+#include "R600Defines.h"
+#include "R600MachineFunctionInfo.h"
+
+using namespace llvm;
+
+R600RegisterInfo::R600RegisterInfo(AMDGPUTargetMachine &tm,
+    const TargetInstrInfo &tii)
+: AMDGPURegisterInfo(tm, tii),
+  TM(tm),
+  TII(tii)
+  { }
+
+BitVector R600RegisterInfo::getReservedRegs(const MachineFunction &MF) const {
+  BitVector Reserved(getNumRegs());
+  const R600MachineFunctionInfo * MFI = MF.getInfo<R600MachineFunctionInfo>();
+
+  Reserved.set(AMDGPU::ZERO);
+  Reserved.set(AMDGPU::HALF);
+  Reserved.set(AMDGPU::ONE);
+  Reserved.set(AMDGPU::ONE_INT);
+  Reserved.set(AMDGPU::NEG_HALF);
+  Reserved.set(AMDGPU::NEG_ONE);
+  Reserved.set(AMDGPU::PV_X);
+  Reserved.set(AMDGPU::ALU_LITERAL_X);
+  Reserved.set(AMDGPU::PREDICATE_BIT);
+  Reserved.set(AMDGPU::PRED_SEL_OFF);
+  Reserved.set(AMDGPU::PRED_SEL_ZERO);
+  Reserved.set(AMDGPU::PRED_SEL_ONE);
+
+  for (TargetRegisterClass::iterator I = AMDGPU::R600_CReg32RegClass.begin(),
+                        E = AMDGPU::R600_CReg32RegClass.end(); I != E; ++I) {
+    Reserved.set(*I);
+  }
+
+  for (std::vector<unsigned>::const_iterator I = MFI->ReservedRegs.begin(),
+                                    E = MFI->ReservedRegs.end(); I != E; ++I) {
+    Reserved.set(*I);
+  }
+
+  return Reserved;
+}
+
+const TargetRegisterClass *
+R600RegisterInfo::getISARegClass(const TargetRegisterClass * rc) const {
+  switch (rc->getID()) {
+  case AMDGPU::GPRF32RegClassID:
+  case AMDGPU::GPRI32RegClassID:
+    return &AMDGPU::R600_Reg32RegClass;
+  default: return rc;
+  }
+}
+
+unsigned R600RegisterInfo::getHWRegChan(unsigned reg) const {
+  return this->getEncodingValue(reg) >> HW_CHAN_SHIFT;
+}
+
+const TargetRegisterClass * R600RegisterInfo::getCFGStructurizerRegClass(
+                                                                   MVT VT) const {
+  switch(VT.SimpleTy) {
+  default:
+  case MVT::i32: return &AMDGPU::R600_TReg32RegClass;
+  }
+}
+
+unsigned R600RegisterInfo::getSubRegFromChannel(unsigned Channel) const {
+  switch (Channel) {
+    default: assert(!"Invalid channel index"); return 0;
+    case 0: return AMDGPU::sel_x;
+    case 1: return AMDGPU::sel_y;
+    case 2: return AMDGPU::sel_z;
+    case 3: return AMDGPU::sel_w;
+  }
+}
diff --git a/lib/Target/R600/R600RegisterInfo.h b/lib/Target/R600/R600RegisterInfo.h
new file mode 100644
index 0000000000..c170ccb378
--- /dev/null
+++ b/lib/Target/R600/R600RegisterInfo.h
@@ -0,0 +1,55 @@
+//===-- R600RegisterInfo.h - R600 Register Info Interface ------*- C++ -*--===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Interface definition for R600RegisterInfo
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef R600REGISTERINFO_H_
+#define R600REGISTERINFO_H_
+
+#include "AMDGPUTargetMachine.h"
+#include "AMDGPURegisterInfo.h"
+
+namespace llvm {
+
+class R600TargetMachine;
+class TargetInstrInfo;
+
+struct R600RegisterInfo : public AMDGPURegisterInfo {
+  AMDGPUTargetMachine &TM;
+  const TargetInstrInfo &TII;
+
+  R600RegisterInfo(AMDGPUTargetMachine &tm, const TargetInstrInfo &tii);
+
+  virtual BitVector getReservedRegs(const MachineFunction &MF) const;
+
+  /// \param RC is an AMDIL reg class.
+  ///
+  /// \returns the R600 reg class that is equivalent to \p RC.
+  virtual const TargetRegisterClass *getISARegClass(
+    const TargetRegisterClass *RC) const;
+
+  /// \brief get the HW encoding for a register's channel.
+  unsigned getHWRegChan(unsigned reg) const;
+
+  /// \brief get the register class of the specified type to use in the
+  /// CFGStructurizer
+  virtual const TargetRegisterClass * getCFGStructurizerRegClass(MVT VT) const;
+
+  /// \returns the sub reg enum value for the given \p Channel
+  /// (e.g. getSubRegFromChannel(0) -> AMDGPU::sel_x)
+  unsigned getSubRegFromChannel(unsigned Channel) const;
+
+};
+
+} // End namespace llvm
+
+#endif // AMDIDSAREGISTERINFO_H_
diff --git a/lib/Target/R600/R600RegisterInfo.td b/lib/Target/R600/R600RegisterInfo.td
new file mode 100644
index 0000000000..d3d6d25d29
--- /dev/null
+++ b/lib/Target/R600/R600RegisterInfo.td
@@ -0,0 +1,107 @@
+
+class R600Reg <string name, bits<16> encoding> : Register<name> {
+  let Namespace = "AMDGPU";
+  let HWEncoding = encoding;
+}
+
+class R600RegWithChan <string name, bits<9> sel, string chan> :
+    Register <name> {
+
+  field bits<2> chan_encoding = !if(!eq(chan, "X"), 0,
+                                !if(!eq(chan, "Y"), 1,
+                                !if(!eq(chan, "Z"), 2,
+                                !if(!eq(chan, "W"), 3, 0))));
+  let HWEncoding{8-0}  = sel;
+  let HWEncoding{10-9} = chan_encoding;
+  let Namespace = "AMDGPU";
+}
+
+class R600Reg_128<string n, list<Register> subregs, bits<16> encoding> :
+    RegisterWithSubRegs<n, subregs> {
+  let Namespace = "AMDGPU";
+  let SubRegIndices = [sel_x, sel_y, sel_z, sel_w];
+  let HWEncoding = encoding;
+}
+
+foreach Index = 0-127 in {
+  foreach Chan = [ "X", "Y", "Z", "W" ] in {
+    // 32-bit Temporary Registers
+    def T#Index#_#Chan : R600RegWithChan <"T"#Index#"."#Chan, Index, Chan>;
+
+    // 32-bit Constant Registers (There are more than 128, this the number
+    // that is currently supported.
+    def C#Index#_#Chan : R600RegWithChan <"C"#Index#"."#Chan, Index, Chan>;
+  }
+  // 128-bit Temporary Registers
+  def T#Index#_XYZW : R600Reg_128 <"T"#Index#".XYZW",
+                                   [!cast<Register>("T"#Index#"_X"),
+                                    !cast<Register>("T"#Index#"_Y"),
+                                    !cast<Register>("T"#Index#"_Z"),
+                                    !cast<Register>("T"#Index#"_W")],
+                                   Index>;
+}
+
+// Array Base Register holding input in FS
+foreach Index = 448-464 in {
+  def ArrayBase#Index :  R600Reg<"ARRAY_BASE", Index>;
+}
+
+
+// Special Registers
+
+def ZERO : R600Reg<"0.0", 248>;
+def ONE : R600Reg<"1.0", 249>;
+def NEG_ONE : R600Reg<"-1.0", 249>;
+def ONE_INT : R600Reg<"1", 250>;
+def HALF : R600Reg<"0.5", 252>;
+def NEG_HALF : R600Reg<"-0.5", 252>;
+def ALU_LITERAL_X : R600Reg<"literal.x", 253>;
+def PV_X : R600Reg<"pv.x", 254>;
+def PREDICATE_BIT : R600Reg<"PredicateBit", 0>;
+def PRED_SEL_OFF: R600Reg<"Pred_sel_off", 0>;
+def PRED_SEL_ZERO : R600Reg<"Pred_sel_zero", 2>;
+def PRED_SEL_ONE : R600Reg<"Pred_sel_one", 3>;
+
+def R600_ArrayBase : RegisterClass <"AMDGPU", [f32, i32], 32,
+                          (add (sequence "ArrayBase%u", 448, 464))>;
+
+def R600_CReg32 : RegisterClass <"AMDGPU", [f32, i32], 32,
+                          (add (interleave
+                                  (interleave (sequence "C%u_X", 0, 127),
+                                              (sequence "C%u_Z", 0, 127)),
+                                  (interleave (sequence "C%u_Y", 0, 127),
+                                              (sequence "C%u_W", 0, 127))))>;
+
+def R600_TReg32_X : RegisterClass <"AMDGPU", [f32, i32], 32,
+                                   (add (sequence "T%u_X", 0, 127))>;
+
+def R600_TReg32_Y : RegisterClass <"AMDGPU", [f32, i32], 32,
+                                   (add (sequence "T%u_Y", 0, 127))>;
+
+def R600_TReg32_Z : RegisterClass <"AMDGPU", [f32, i32], 32,
+                                   (add (sequence "T%u_Z", 0, 127))>;
+
+def R600_TReg32_W : RegisterClass <"AMDGPU", [f32, i32], 32,
+                                   (add (sequence "T%u_W", 0, 127))>;
+
+def R600_TReg32 : RegisterClass <"AMDGPU", [f32, i32], 32,
+                          (add (interleave
+                                 (interleave R600_TReg32_X, R600_TReg32_Z),
+                                 (interleave R600_TReg32_Y, R600_TReg32_W)))>;
+
+def R600_Reg32 : RegisterClass <"AMDGPU", [f32, i32], 32, (add
+    R600_TReg32,
+    R600_CReg32,
+    R600_ArrayBase,
+    ZERO, HALF, ONE, ONE_INT, PV_X, ALU_LITERAL_X, NEG_ONE, NEG_HALF)>;
+
+def R600_Predicate : RegisterClass <"AMDGPU", [i32], 32, (add
+    PRED_SEL_OFF, PRED_SEL_ZERO, PRED_SEL_ONE)>;
+
+def R600_Predicate_Bit: RegisterClass <"AMDGPU", [i32], 32, (add
+    PREDICATE_BIT)>;
+
+def R600_Reg128 : RegisterClass<"AMDGPU", [v4f32, v4i32], 128,
+                                (add (sequence "T%u_XYZW", 0, 127))> {
+  let CopyCost = -1;
+}
diff --git a/lib/Target/R600/R600Schedule.td b/lib/Target/R600/R600Schedule.td
new file mode 100644
index 0000000000..7ede181c51
--- /dev/null
+++ b/lib/Target/R600/R600Schedule.td
@@ -0,0 +1,36 @@
+//===-- R600Schedule.td - R600 Scheduling definitions ------*- tablegen -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// R600 has a VLIW architecture.  On pre-cayman cards there are 5 instruction
+// slots ALU.X, ALU.Y, ALU.Z, ALU.W, and TRANS.  For cayman cards, the TRANS
+// slot has been removed. 
+//
+//===----------------------------------------------------------------------===//
+
+
+def ALU_X : FuncUnit;
+def ALU_Y : FuncUnit;
+def ALU_Z : FuncUnit;
+def ALU_W : FuncUnit;
+def TRANS : FuncUnit;
+
+def AnyALU : InstrItinClass;
+def VecALU : InstrItinClass;
+def TransALU : InstrItinClass;
+
+def R600_EG_Itin : ProcessorItineraries <
+  [ALU_X, ALU_Y, ALU_Z, ALU_W, TRANS, ALU_NULL],
+  [],
+  [
+    InstrItinData<AnyALU, [InstrStage<1, [ALU_X, ALU_Y, ALU_Z, ALU_W, TRANS]>]>,
+    InstrItinData<VecALU, [InstrStage<1, [ALU_X, ALU_Y, ALU_X, ALU_W]>]>,
+    InstrItinData<TransALU, [InstrStage<1, [TRANS]>]>,
+    InstrItinData<NullALU, [InstrStage<1, [ALU_NULL]>]>
+  ]
+>;
diff --git a/lib/Target/R600/SIAssignInterpRegs.cpp b/lib/Target/R600/SIAssignInterpRegs.cpp
new file mode 100644
index 0000000000..832e44d766
--- /dev/null
+++ b/lib/Target/R600/SIAssignInterpRegs.cpp
@@ -0,0 +1,152 @@
+//===-- SIAssignInterpRegs.cpp - Assign interpolation registers -----------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief This pass maps the pseudo interpolation registers to the correct physical
+/// registers.
+//
+/// Prior to executing a fragment shader, the GPU loads interpolation
+/// parameters into physical registers.  The specific physical register that each
+/// interpolation parameter ends up in depends on the type of the interpolation
+/// parameter as well as how many interpolation parameters are used by the
+/// shader.
+//
+//===----------------------------------------------------------------------===//
+
+
+
+#include "AMDGPU.h"
+#include "AMDIL.h"
+#include "SIMachineFunctionInfo.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+
+using namespace llvm;
+
+namespace {
+
+class SIAssignInterpRegsPass : public MachineFunctionPass {
+
+private:
+  static char ID;
+  TargetMachine &TM;
+
+  void addLiveIn(MachineFunction * MF,  MachineRegisterInfo & MRI,
+                 unsigned physReg, unsigned virtReg);
+
+public:
+  SIAssignInterpRegsPass(TargetMachine &tm) :
+    MachineFunctionPass(ID), TM(tm) { }
+
+  virtual bool runOnMachineFunction(MachineFunction &MF);
+
+  const char *getPassName() const { return "SI Assign intrpolation registers"; }
+};
+
+} // End anonymous namespace
+
+char SIAssignInterpRegsPass::ID = 0;
+
+#define INTERP_VALUES 16
+#define REQUIRED_VALUE_MAX_INDEX 7
+
+struct InterpInfo {
+  bool Enabled;
+  unsigned Regs[3];
+  unsigned RegCount;
+};
+
+
+FunctionPass *llvm::createSIAssignInterpRegsPass(TargetMachine &tm) {
+  return new SIAssignInterpRegsPass(tm);
+}
+
+bool SIAssignInterpRegsPass::runOnMachineFunction(MachineFunction &MF) {
+
+  struct InterpInfo InterpUse[INTERP_VALUES] = {
+    {false, {AMDGPU::PERSP_SAMPLE_I, AMDGPU::PERSP_SAMPLE_J}, 2},
+    {false, {AMDGPU::PERSP_CENTER_I, AMDGPU::PERSP_CENTER_J}, 2},
+    {false, {AMDGPU::PERSP_CENTROID_I, AMDGPU::PERSP_CENTROID_J}, 2},
+    {false, {AMDGPU::PERSP_I_W, AMDGPU::PERSP_J_W, AMDGPU::PERSP_1_W}, 3},
+    {false, {AMDGPU::LINEAR_SAMPLE_I, AMDGPU::LINEAR_SAMPLE_J}, 2},
+    {false, {AMDGPU::LINEAR_CENTER_I, AMDGPU::LINEAR_CENTER_J}, 2},
+    {false, {AMDGPU::LINEAR_CENTROID_I, AMDGPU::LINEAR_CENTROID_J}, 2},
+    {false, {AMDGPU::LINE_STIPPLE_TEX_COORD}, 1},
+    {false, {AMDGPU::POS_X_FLOAT}, 1},
+    {false, {AMDGPU::POS_Y_FLOAT}, 1},
+    {false, {AMDGPU::POS_Z_FLOAT}, 1},
+    {false, {AMDGPU::POS_W_FLOAT}, 1},
+    {false, {AMDGPU::FRONT_FACE}, 1},
+    {false, {AMDGPU::ANCILLARY}, 1},
+    {false, {AMDGPU::SAMPLE_COVERAGE}, 1},
+    {false, {AMDGPU::POS_FIXED_PT}, 1}
+  };
+
+  SIMachineFunctionInfo * MFI = MF.getInfo<SIMachineFunctionInfo>();
+  // This pass is only needed for pixel shaders.
+  if (MFI->ShaderType != ShaderType::PIXEL) {
+    return false;
+  }
+  MachineRegisterInfo &MRI = MF.getRegInfo();
+  bool ForceEnable = true;
+
+  // First pass, mark the interpolation values that are used.
+  for (unsigned InterpIdx = 0; InterpIdx < INTERP_VALUES; InterpIdx++) {
+    for (unsigned RegIdx = 0; RegIdx < InterpUse[InterpIdx].RegCount;
+                                                               RegIdx++) {
+      InterpUse[InterpIdx].Enabled = InterpUse[InterpIdx].Enabled ||
+                            !MRI.use_empty(InterpUse[InterpIdx].Regs[RegIdx]);
+      if (InterpUse[InterpIdx].Enabled &&
+          InterpIdx <= REQUIRED_VALUE_MAX_INDEX) {
+        ForceEnable = false;
+      }
+    }
+  }
+
+  // At least one interpolation mode must be enabled or else the GPU will hang.
+  if (ForceEnable) {
+    InterpUse[0].Enabled = true;
+  }
+
+  unsigned UsedVgprs = 0;
+
+  // Second pass, replace with VGPRs.
+  for (unsigned InterpIdx = 0; InterpIdx < INTERP_VALUES; InterpIdx++) {
+    if (!InterpUse[InterpIdx].Enabled) {
+      continue;
+    }
+    MFI->SPIPSInputAddr |= (1 << InterpIdx);
+
+    for (unsigned RegIdx = 0; RegIdx < InterpUse[InterpIdx].RegCount;
+                                                  RegIdx++, UsedVgprs++) {
+      unsigned NewReg = AMDGPU::VReg_32RegClass.getRegister(UsedVgprs);
+      unsigned VirtReg = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
+      MRI.replaceRegWith(InterpUse[InterpIdx].Regs[RegIdx], VirtReg);
+      addLiveIn(&MF, MRI, NewReg, VirtReg);
+    }
+  }
+
+  return false;
+}
+
+void SIAssignInterpRegsPass::addLiveIn(MachineFunction * MF,
+                           MachineRegisterInfo & MRI,
+                           unsigned physReg, unsigned virtReg) {
+    const TargetInstrInfo * TII = TM.getInstrInfo();
+    if (!MRI.isLiveIn(physReg)) {
+      MRI.addLiveIn(physReg, virtReg);
+      MF->front().addLiveIn(physReg);
+      BuildMI(MF->front(), MF->front().begin(), DebugLoc(),
+              TII->get(TargetOpcode::COPY), virtReg)
+                .addReg(physReg);
+    } else {
+      MRI.replaceRegWith(virtReg, MRI.getLiveInVirtReg(physReg));
+    }
+}
diff --git a/lib/Target/R600/SIFixSGPRLiveness.cpp b/lib/Target/R600/SIFixSGPRLiveness.cpp
new file mode 100644
index 0000000000..0fecd7a28e
--- /dev/null
+++ b/lib/Target/R600/SIFixSGPRLiveness.cpp
@@ -0,0 +1,179 @@
+//===-- SIFixSGPRLiveness.cpp - SGPR liveness adjustment ------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+///
+/// SGPRs are not affected by control flow. This pass adjusts SGPR liveness in
+/// so that the register allocator can still correctly allocate them.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/MachineDominators.h"
+#include "llvm/CodeGen/MachinePostDominators.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+
+using namespace llvm;
+
+namespace {
+
+class SIFixSGPRLiveness : public MachineFunctionPass {
+private:
+  static char ID;
+
+  const TargetInstrInfo *TII;
+  MachineRegisterInfo *MRI;
+  MachineDominatorTree *MD;
+  MachinePostDominatorTree *MPD;
+
+  bool isSGPR(const TargetRegisterClass *RegClass) {
+    return RegClass == &AMDGPU::SReg_1RegClass ||
+           RegClass == &AMDGPU::SReg_32RegClass ||
+           RegClass == &AMDGPU::SReg_64RegClass ||
+           RegClass == &AMDGPU::SReg_128RegClass ||
+           RegClass == &AMDGPU::SReg_256RegClass;
+  }
+
+  void addKill(MachineBasicBlock::iterator I, unsigned Reg);
+  MachineBasicBlock *handleUses(unsigned VirtReg, MachineBasicBlock *Begin);
+  void handlePreds(MachineBasicBlock *Begin, MachineBasicBlock *End,
+                   unsigned VirtReg);
+
+  bool handleVirtReg(unsigned VirtReg);
+
+public:
+  SIFixSGPRLiveness(TargetMachine &tm);
+
+  virtual bool runOnMachineFunction(MachineFunction &MF);
+
+  virtual const char *getPassName() const {
+    return "SI fix SGPR liveness pass";
+  }
+
+  virtual void getAnalysisUsage(AnalysisUsage &AU) const;
+};
+
+} // end anonymous namespace
+
+char SIFixSGPRLiveness::ID = 0;
+
+SIFixSGPRLiveness::SIFixSGPRLiveness(TargetMachine &tm):
+  MachineFunctionPass(ID),
+  TII(tm.getInstrInfo()) {
+  initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
+}
+
+void SIFixSGPRLiveness::getAnalysisUsage(AnalysisUsage &AU) const {
+  AU.addRequired<MachineDominatorTree>();
+  AU.addRequired<MachinePostDominatorTree>();
+  AU.setPreservesCFG();
+  MachineFunctionPass::getAnalysisUsage(AU);
+}
+
+void SIFixSGPRLiveness::addKill(MachineBasicBlock::iterator I, unsigned Reg) {
+  MachineBasicBlock *MBB = I->getParent();
+
+  BuildMI(*MBB, I, DebugLoc(), TII->get(TargetOpcode::KILL)).addReg(Reg);
+}
+
+// Find the common post dominator of all uses
+MachineBasicBlock *SIFixSGPRLiveness::handleUses(unsigned VirtReg,
+                                                 MachineBasicBlock *Begin) {
+  MachineBasicBlock *LastUse = Begin, *End = Begin;
+  bool EndUsesReg = true;
+
+  MachineRegisterInfo::use_iterator i, e;
+  for (i = MRI->use_begin(VirtReg), e = MRI->use_end(); i != e; ++i) {
+    MachineBasicBlock *MBB = i->getParent();
+    if (LastUse == MBB)
+      continue;
+
+    LastUse = MBB;
+    MBB = MPD->findNearestCommonDominator(End, MBB);
+
+    if (MBB == LastUse)
+      EndUsesReg = true;
+    else if (MBB != End)
+      EndUsesReg = false;
+
+    End = MBB;
+  }
+
+  return EndUsesReg ? Begin : End;
+}
+
+// Handles predecessors separately, only add KILLs to dominated ones
+void SIFixSGPRLiveness::handlePreds(MachineBasicBlock *Begin,
+                                    MachineBasicBlock *End,
+                                    unsigned VirtReg) {
+  MachineBasicBlock::pred_iterator i, e;
+  for (i = End->pred_begin(), e = End->pred_end(); i != e; ++i) {
+
+    if (MD->dominates(End, *i))
+      continue; // ignore loops
+
+    if (MD->dominates(*i, Begin))
+      continue; // too far up, abort search
+
+    if (MD->dominates(Begin, *i)) {
+      // found end of livetime
+      addKill((*i)->getFirstTerminator(), VirtReg);
+      continue;
+    }
+
+    handlePreds(Begin, *i, VirtReg);
+  }
+}
+
+bool SIFixSGPRLiveness::handleVirtReg(unsigned VirtReg) {
+
+  MachineInstr *Def = MRI->getVRegDef(VirtReg);
+  if (!Def || MRI->use_empty(VirtReg))
+    return false; // No definition or not used
+
+  MachineBasicBlock *Begin = Def->getParent();
+  MachineBasicBlock *End = handleUses(VirtReg, Begin);
+  if (Begin == End)
+    return false; // Defined and only used in the same block
+
+  if (MD->dominates(Begin, End)) {
+    // Lifetime dominate the end node, just kill it here
+    addKill(End->getFirstNonPHI(), VirtReg);
+  } else {
+    // only some predecessors are dominate, handle them separately
+    handlePreds(Begin, End, VirtReg);
+  }
+  return true;
+}
+
+bool SIFixSGPRLiveness::runOnMachineFunction(MachineFunction &MF) {
+  bool Changes = false;
+
+  MRI = &MF.getRegInfo();
+  MD = &getAnalysis<MachineDominatorTree>();
+  MPD = &getAnalysis<MachinePostDominatorTree>();
+
+  for (unsigned i = 0, e = MRI->getNumVirtRegs(); i != e; ++i) {
+    unsigned VirtReg = TargetRegisterInfo::index2VirtReg(i);
+
+    const TargetRegisterClass *RegClass = MRI->getRegClass(VirtReg);
+    if (!isSGPR(RegClass))
+      continue;
+
+    Changes |= handleVirtReg(VirtReg);
+  }
+
+  return Changes;
+}
+
+FunctionPass *llvm::createSIFixSGPRLivenessPass(TargetMachine &tm) {
+  return new SIFixSGPRLiveness(tm);
+}
diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
new file mode 100644
index 0000000000..292ce850f7
--- /dev/null
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -0,0 +1,442 @@
+//===-- SIISelLowering.cpp - SI DAG Lowering Implementation ---------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Custom DAG lowering for SI
+//
+//===----------------------------------------------------------------------===//
+
+#include "SIISelLowering.h"
+#include "AMDIL.h"
+#include "AMDILIntrinsicInfo.h"
+#include "SIInstrInfo.h"
+#include "SIMachineFunctionInfo.h"
+#include "SIRegisterInfo.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/SelectionDAG.h"
+
+using namespace llvm;
+
+SITargetLowering::SITargetLowering(TargetMachine &TM) :
+    AMDGPUTargetLowering(TM),
+    TII(static_cast<const SIInstrInfo*>(TM.getInstrInfo())) {
+  addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
+  addRegisterClass(MVT::f32, &AMDGPU::VReg_32RegClass);
+  addRegisterClass(MVT::i32, &AMDGPU::VReg_32RegClass);
+  addRegisterClass(MVT::i64, &AMDGPU::SReg_64RegClass);
+  addRegisterClass(MVT::i1, &AMDGPU::SCCRegRegClass);
+  addRegisterClass(MVT::i1, &AMDGPU::VCCRegRegClass);
+
+  addRegisterClass(MVT::v4i32, &AMDGPU::SReg_128RegClass);
+  addRegisterClass(MVT::v8i32, &AMDGPU::SReg_256RegClass);
+
+  computeRegisterProperties();
+
+  setOperationAction(ISD::AND, MVT::i1, Custom);
+
+  setOperationAction(ISD::ADD, MVT::i64, Legal);
+  setOperationAction(ISD::ADD, MVT::i32, Legal);
+
+  setOperationAction(ISD::BR_CC, MVT::i32, Custom);
+
+  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
+
+  // We need to custom lower loads from the USER_SGPR address space, so we can
+  // add the SGPRs as livein registers.
+  setOperationAction(ISD::LOAD, MVT::i32, Custom);
+  setOperationAction(ISD::LOAD, MVT::i64, Custom);
+
+  setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
+  setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
+
+  setOperationAction(ISD::SELECT_CC, MVT::Other, Expand);
+  setTargetDAGCombine(ISD::SELECT_CC);
+
+  setTargetDAGCombine(ISD::SETCC);
+}
+
+MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
+    MachineInstr * MI, MachineBasicBlock * BB) const {
+  const TargetInstrInfo * TII = getTargetMachine().getInstrInfo();
+  MachineRegisterInfo & MRI = BB->getParent()->getRegInfo();
+  MachineBasicBlock::iterator I = MI;
+
+  if (TII->get(MI->getOpcode()).TSFlags & SIInstrFlags::NEED_WAIT) {
+    AppendS_WAITCNT(MI, *BB, llvm::next(I));
+    return BB;
+  }
+
+  switch (MI->getOpcode()) {
+  default:
+    return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
+  case AMDGPU::BRANCH: return BB;
+  case AMDGPU::CLAMP_SI:
+    BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::V_MOV_B32_e64))
+           .addOperand(MI->getOperand(0))
+           .addOperand(MI->getOperand(1))
+           // VSRC1-2 are unused, but we still need to fill all the
+           // operand slots, so we just reuse the VSRC0 operand
+           .addOperand(MI->getOperand(1))
+           .addOperand(MI->getOperand(1))
+           .addImm(0) // ABS
+           .addImm(1) // CLAMP
+           .addImm(0) // OMOD
+           .addImm(0); // NEG
+    MI->eraseFromParent();
+    break;
+
+  case AMDGPU::FABS_SI:
+    BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::V_MOV_B32_e64))
+                 .addOperand(MI->getOperand(0))
+                 .addOperand(MI->getOperand(1))
+                 // VSRC1-2 are unused, but we still need to fill all the
+                 // operand slots, so we just reuse the VSRC0 operand
+                 .addOperand(MI->getOperand(1))
+                 .addOperand(MI->getOperand(1))
+                 .addImm(1) // ABS
+                 .addImm(0) // CLAMP
+                 .addImm(0) // OMOD
+                 .addImm(0); // NEG
+    MI->eraseFromParent();
+    break;
+
+  case AMDGPU::FNEG_SI:
+    BuildMI(*BB, I, BB->findDebugLoc(I), TII->get(AMDGPU::V_MOV_B32_e64))
+                 .addOperand(MI->getOperand(0))
+                 .addOperand(MI->getOperand(1))
+                 // VSRC1-2 are unused, but we still need to fill all the
+                 // operand slots, so we just reuse the VSRC0 operand
+                 .addOperand(MI->getOperand(1))
+                 .addOperand(MI->getOperand(1))
+                 .addImm(0) // ABS
+                 .addImm(0) // CLAMP
+                 .addImm(0) // OMOD
+                 .addImm(1); // NEG
+    MI->eraseFromParent();
+    break;
+  case AMDGPU::SHADER_TYPE:
+    BB->getParent()->getInfo<SIMachineFunctionInfo>()->ShaderType =
+                                        MI->getOperand(0).getImm();
+    MI->eraseFromParent();
+    break;
+
+  case AMDGPU::SI_INTERP:
+    LowerSI_INTERP(MI, *BB, I, MRI);
+    break;
+  case AMDGPU::SI_INTERP_CONST:
+    LowerSI_INTERP_CONST(MI, *BB, I, MRI);
+    break;
+  case AMDGPU::SI_KIL:
+    LowerSI_KIL(MI, *BB, I, MRI);
+    break;
+  case AMDGPU::SI_WQM:
+    LowerSI_WQM(MI, *BB, I, MRI);
+    break;
+  case AMDGPU::SI_V_CNDLT:
+    LowerSI_V_CNDLT(MI, *BB, I, MRI);
+    break;
+  }
+  return BB;
+}
+
+void SITargetLowering::AppendS_WAITCNT(MachineInstr *MI, MachineBasicBlock &BB,
+    MachineBasicBlock::iterator I) const {
+  BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_WAITCNT))
+          .addImm(0);
+}
+
+
+void SITargetLowering::LowerSI_WQM(MachineInstr *MI, MachineBasicBlock &BB,
+    MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const {
+  BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_WQM_B64), AMDGPU::EXEC)
+          .addReg(AMDGPU::EXEC);
+
+  MI->eraseFromParent();
+}
+
+void SITargetLowering::LowerSI_INTERP(MachineInstr *MI, MachineBasicBlock &BB,
+    MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const {
+  unsigned tmp = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
+  unsigned M0 = MRI.createVirtualRegister(&AMDGPU::M0RegRegClass);
+  MachineOperand dst = MI->getOperand(0);
+  MachineOperand iReg = MI->getOperand(1);
+  MachineOperand jReg = MI->getOperand(2);
+  MachineOperand attr_chan = MI->getOperand(3);
+  MachineOperand attr = MI->getOperand(4);
+  MachineOperand params = MI->getOperand(5);
+
+  BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_MOV_B32), M0)
+          .addOperand(params);
+
+  BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_INTERP_P1_F32), tmp)
+          .addOperand(iReg)
+          .addOperand(attr_chan)
+          .addOperand(attr)
+          .addReg(M0);
+
+  BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_INTERP_P2_F32))
+          .addOperand(dst)
+          .addReg(tmp)
+          .addOperand(jReg)
+          .addOperand(attr_chan)
+          .addOperand(attr)
+          .addReg(M0);
+
+  MI->eraseFromParent();
+}
+
+void SITargetLowering::LowerSI_INTERP_CONST(MachineInstr *MI,
+    MachineBasicBlock &BB, MachineBasicBlock::iterator I,
+    MachineRegisterInfo &MRI) const {
+  MachineOperand dst = MI->getOperand(0);
+  MachineOperand attr_chan = MI->getOperand(1);
+  MachineOperand attr = MI->getOperand(2);
+  MachineOperand params = MI->getOperand(3);
+  unsigned M0 = MRI.createVirtualRegister(&AMDGPU::M0RegRegClass);
+
+  BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::S_MOV_B32), M0)
+          .addOperand(params);
+
+  BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_INTERP_MOV_F32))
+          .addOperand(dst)
+          .addOperand(attr_chan)
+          .addOperand(attr)
+          .addReg(M0);
+
+  MI->eraseFromParent();
+}
+
+void SITargetLowering::LowerSI_KIL(MachineInstr *MI, MachineBasicBlock &BB,
+    MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const {
+  // Clear this pixel from the exec mask if the operand is negative
+  BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_CMPX_LE_F32_e32),
+          AMDGPU::VCC)
+          .addReg(AMDGPU::SREG_LIT_0)
+          .addOperand(MI->getOperand(0));
+
+  MI->eraseFromParent();
+}
+
+void SITargetLowering::LowerSI_V_CNDLT(MachineInstr *MI, MachineBasicBlock &BB,
+    MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const {
+  unsigned VCC = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
+
+  BuildMI(BB, I, BB.findDebugLoc(I),
+          TII->get(AMDGPU::V_CMP_GT_F32_e32),
+          VCC)
+          .addReg(AMDGPU::SREG_LIT_0)
+          .addOperand(MI->getOperand(1));
+
+  BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_CNDMASK_B32_e32))
+          .addOperand(MI->getOperand(0))
+          .addOperand(MI->getOperand(3))
+          .addOperand(MI->getOperand(2))
+          .addReg(VCC);
+
+  MI->eraseFromParent();
+}
+
+EVT SITargetLowering::getSetCCResultType(EVT VT) const {
+  return MVT::i1;
+}
+
+//===----------------------------------------------------------------------===//
+// Custom DAG Lowering Operations
+//===----------------------------------------------------------------------===//
+
+SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
+  switch (Op.getOpcode()) {
+  default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
+  case ISD::BR_CC: return LowerBR_CC(Op, DAG);
+  case ISD::LOAD: return LowerLOAD(Op, DAG);
+  case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
+  case ISD::AND: return Loweri1ContextSwitch(Op, DAG, ISD::AND);
+  case ISD::INTRINSIC_WO_CHAIN: {
+    unsigned IntrinsicID =
+                         cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
+    EVT VT = Op.getValueType();
+    switch (IntrinsicID) {
+    case AMDGPUIntrinsic::SI_vs_load_buffer_index:
+      return CreateLiveInRegister(DAG, &AMDGPU::VReg_32RegClass,
+                                  AMDGPU::VGPR0, VT);
+    default: return AMDGPUTargetLowering::LowerOperation(Op, DAG);
+    }
+    break;
+  }
+  }
+  return SDValue();
+}
+
+/// \brief The function is for lowering i1 operations on the
+/// VCC register.
+///
+/// In the VALU context, VCC is a one bit register, but in the
+/// SALU context the VCC is a 64-bit register (1-bit per thread).  Since only
+/// the SALU can perform operations on the VCC register, we need to promote
+/// the operand types from i1 to i64 in order for tablegen to be able to match
+/// this operation to the correct SALU instruction.  We do this promotion by
+/// wrapping the operands in a CopyToReg node.
+///
+SDValue SITargetLowering::Loweri1ContextSwitch(SDValue Op,
+                                               SelectionDAG &DAG,
+                                               unsigned VCCNode) const {
+  DebugLoc DL = Op.getDebugLoc();
+
+  SDValue OpNode = DAG.getNode(VCCNode, DL, MVT::i64,
+                               DAG.getNode(SIISD::VCC_BITCAST, DL, MVT::i64,
+                                           Op.getOperand(0)),
+                               DAG.getNode(SIISD::VCC_BITCAST, DL, MVT::i64,
+                                           Op.getOperand(1)));
+
+  return DAG.getNode(SIISD::VCC_BITCAST, DL, MVT::i1, OpNode);
+}
+
+SDValue SITargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
+  SDValue Chain = Op.getOperand(0);
+  SDValue CC = Op.getOperand(1);
+  SDValue LHS   = Op.getOperand(2);
+  SDValue RHS   = Op.getOperand(3);
+  SDValue JumpT  = Op.getOperand(4);
+  SDValue CmpValue;
+  SDValue Result;
+  CmpValue = DAG.getNode(
+      ISD::SETCC,
+      Op.getDebugLoc(),
+      MVT::i1,
+      LHS, RHS,
+      CC);
+
+  Result = DAG.getNode(
+      AMDGPUISD::BRANCH_COND,
+      CmpValue.getDebugLoc(),
+      MVT::Other, Chain,
+      JumpT, CmpValue);
+  return Result;
+}
+
+SDValue SITargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
+  EVT VT = Op.getValueType();
+  LoadSDNode *Ptr = dyn_cast<LoadSDNode>(Op);
+
+  assert(Ptr);
+
+  unsigned AddrSpace = Ptr->getPointerInfo().getAddrSpace();
+
+  // We only need to lower USER_SGPR address space loads
+  if (AddrSpace != AMDGPUAS::USER_SGPR_ADDRESS) {
+    return SDValue();
+  }
+
+  // Loads from the USER_SGPR address space can only have constant value
+  // pointers.
+  ConstantSDNode *BasePtr = dyn_cast<ConstantSDNode>(Ptr->getBasePtr());
+  assert(BasePtr);
+
+  unsigned TypeDwordWidth = VT.getSizeInBits() / 32;
+  const TargetRegisterClass * dstClass;
+  switch (TypeDwordWidth) {
+    default:
+      assert(!"USER_SGPR value size not implemented");
+      return SDValue();
+    case 1:
+      dstClass = &AMDGPU::SReg_32RegClass;
+      break;
+    case 2:
+      dstClass = &AMDGPU::SReg_64RegClass;
+      break;
+  }
+  uint64_t Index = BasePtr->getZExtValue();
+  assert(Index % TypeDwordWidth == 0 && "USER_SGPR not properly aligned");
+  unsigned SGPRIndex = Index / TypeDwordWidth;
+  unsigned Reg = dstClass->getRegister(SGPRIndex);
+
+  DAG.ReplaceAllUsesOfValueWith(Op, CreateLiveInRegister(DAG, dstClass, Reg,
+                                                         VT));
+  return SDValue();
+}
+
+SDValue SITargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
+  SDValue LHS = Op.getOperand(0);
+  SDValue RHS = Op.getOperand(1);
+  SDValue True = Op.getOperand(2);
+  SDValue False = Op.getOperand(3);
+  SDValue CC = Op.getOperand(4);
+  EVT VT = Op.getValueType();
+  DebugLoc DL = Op.getDebugLoc();
+
+  // Possible Min/Max pattern
+  SDValue MinMax = LowerMinMax(Op, DAG);
+  if (MinMax.getNode()) {
+    return MinMax;
+  }
+
+  SDValue Cond = DAG.getNode(ISD::SETCC, DL, MVT::i1, LHS, RHS, CC);
+  return DAG.getNode(ISD::SELECT, DL, VT, Cond, True, False);
+}
+
+//===----------------------------------------------------------------------===//
+// Custom DAG optimizations
+//===----------------------------------------------------------------------===//
+
+SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
+                                            DAGCombinerInfo &DCI) const {
+  SelectionDAG &DAG = DCI.DAG;
+  DebugLoc DL = N->getDebugLoc();
+  EVT VT = N->getValueType(0);
+
+  switch (N->getOpcode()) {
+    default: break;
+    case ISD::SELECT_CC: {
+      N->dump();
+      ConstantSDNode *True, *False;
+      // i1 selectcc(l, r, -1, 0, cc) -> i1 setcc(l, r, cc)
+      if ((True = dyn_cast<ConstantSDNode>(N->getOperand(2)))
+          && (False = dyn_cast<ConstantSDNode>(N->getOperand(3)))
+          && True->isAllOnesValue()
+          && False->isNullValue()
+          && VT == MVT::i1) {
+        return DAG.getNode(ISD::SETCC, DL, VT, N->getOperand(0),
+                           N->getOperand(1), N->getOperand(4));
+
+      }
+      break;
+    }
+    case ISD::SETCC: {
+      SDValue Arg0 = N->getOperand(0);
+      SDValue Arg1 = N->getOperand(1);
+      SDValue CC = N->getOperand(2);
+      ConstantSDNode * C = NULL;
+      ISD::CondCode CCOp = dyn_cast<CondCodeSDNode>(CC)->get();
+
+      // i1 setcc (sext(i1), 0, setne) -> i1 setcc(i1, 0, setne)
+      if (VT == MVT::i1
+          && Arg0.getOpcode() == ISD::SIGN_EXTEND
+          && Arg0.getOperand(0).getValueType() == MVT::i1
+          && (C = dyn_cast<ConstantSDNode>(Arg1))
+          && C->isNullValue()
+          && CCOp == ISD::SETNE) {
+        return SimplifySetCC(VT, Arg0.getOperand(0),
+                             DAG.getConstant(0, MVT::i1), CCOp, true, DCI, DL);
+      }
+      break;
+    }
+  }
+  return SDValue();
+}
+
+#define NODE_NAME_CASE(node) case SIISD::node: return #node;
+
+const char* SITargetLowering::getTargetNodeName(unsigned Opcode) const {
+  switch (Opcode) {
+  default: return AMDGPUTargetLowering::getTargetNodeName(Opcode);
+  NODE_NAME_CASE(VCC_AND)
+  NODE_NAME_CASE(VCC_BITCAST)
+  }
+}
diff --git a/lib/Target/R600/SIISelLowering.h b/lib/Target/R600/SIISelLowering.h
new file mode 100644
index 0000000000..27c2a1c39a
--- /dev/null
+++ b/lib/Target/R600/SIISelLowering.h
@@ -0,0 +1,62 @@
+//===-- SIISelLowering.h - SI DAG Lowering Interface ------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief SI DAG Lowering interface definition
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef SIISELLOWERING_H
+#define SIISELLOWERING_H
+
+#include "AMDGPUISelLowering.h"
+#include "SIInstrInfo.h"
+
+namespace llvm {
+
+class SITargetLowering : public AMDGPUTargetLowering {
+  const SIInstrInfo * TII;
+
+  /// Memory reads and writes are syncronized using the S_WAITCNT instruction.
+  /// This function takes the most conservative approach and inserts an
+  /// S_WAITCNT instruction after every read and write.
+  void AppendS_WAITCNT(MachineInstr *MI, MachineBasicBlock &BB,
+              MachineBasicBlock::iterator I) const;
+  void LowerMOV_IMM(MachineInstr *MI, MachineBasicBlock &BB,
+              MachineBasicBlock::iterator I, unsigned Opocde) const;
+  void LowerSI_INTERP(MachineInstr *MI, MachineBasicBlock &BB,
+              MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
+  void LowerSI_INTERP_CONST(MachineInstr *MI, MachineBasicBlock &BB,
+              MachineBasicBlock::iterator I, MachineRegisterInfo &MRI) const;
+  void LowerSI_KIL(MachineInstr *MI, MachineBasicBlock &BB,
+              MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
+  void LowerSI_WQM(MachineInstr *MI, MachineBasicBlock &BB,
+              MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
+  void LowerSI_V_CNDLT(MachineInstr *MI, MachineBasicBlock &BB,
+              MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
+
+  SDValue Loweri1ContextSwitch(SDValue Op, SelectionDAG &DAG,
+                                           unsigned VCCNode) const;
+  SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
+
+public:
+  SITargetLowering(TargetMachine &tm);
+  virtual MachineBasicBlock * EmitInstrWithCustomInserter(MachineInstr * MI,
+                                              MachineBasicBlock * BB) const;
+  virtual EVT getSetCCResultType(EVT VT) const;
+  virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
+  virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
+  virtual const char* getTargetNodeName(unsigned Opcode) const;
+};
+
+} // End namespace llvm
+
+#endif //SIISELLOWERING_H
diff --git a/lib/Target/R600/SIInstrFormats.td b/lib/Target/R600/SIInstrFormats.td
new file mode 100644
index 0000000000..aea3b5a888
--- /dev/null
+++ b/lib/Target/R600/SIInstrFormats.td
@@ -0,0 +1,146 @@
+//===-- SIInstrFormats.td - SI Instruction Formats ------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// SI Instruction format definitions.
+//
+// Instructions with _32 take 32-bit operands.
+// Instructions with _64 take 64-bit operands.
+//
+// VOP_* instructions can use either a 32-bit or 64-bit encoding.  The 32-bit
+// encoding is the standard encoding, but instruction that make use of
+// any of the instruction modifiers must use the 64-bit encoding.
+//
+// Instructions with _e32 use the 32-bit encoding.
+// Instructions with _e64 use the 64-bit encoding.
+//
+//===----------------------------------------------------------------------===//
+
+class VOP3b_2IN <bits<9> op, string opName, RegisterClass dstClass,
+                 RegisterClass src0Class, RegisterClass src1Class,
+                 list<dag> pattern>
+  : VOP3b <op, (outs dstClass:$vdst),
+               (ins src0Class:$src0, src1Class:$src1, InstFlag:$src2, InstFlag:$sdst,
+                    InstFlag:$omod, InstFlag:$neg),
+           opName, pattern
+>;
+
+
+class VOP3_1_32 <bits<9> op, string opName, list<dag> pattern>
+  : VOP3b_2IN <op, opName, SReg_1, AllReg_32, VReg_32, pattern>;
+
+class VOP3_32 <bits<9> op, string opName, list<dag> pattern>
+  : VOP3 <op, (outs VReg_32:$dst), (ins AllReg_32:$src0, VReg_32:$src1, VReg_32:$src2, i32imm:$src3, i32imm:$src4, i32imm:$src5, i32imm:$src6), opName, pattern>;
+
+class VOP3_64 <bits<9> op, string opName, list<dag> pattern>
+  : VOP3 <op, (outs VReg_64:$dst), (ins AllReg_64:$src0, VReg_64:$src1, VReg_64:$src2, i32imm:$src3, i32imm:$src4, i32imm:$src5, i32imm:$src6), opName, pattern>;
+
+
+class SOP1_32 <bits<8> op, string opName, list<dag> pattern>
+  : SOP1 <op, (outs SReg_32:$dst), (ins SReg_32:$src0), opName, pattern>;
+
+class SOP1_64 <bits<8> op, string opName, list<dag> pattern>
+  : SOP1 <op, (outs SReg_64:$dst), (ins SReg_64:$src0), opName, pattern>;
+
+class SOP2_32 <bits<7> op, string opName, list<dag> pattern>
+  : SOP2 <op, (outs SReg_32:$dst), (ins SReg_32:$src0, SReg_32:$src1), opName, pattern>;
+
+class SOP2_64 <bits<7> op, string opName, list<dag> pattern>
+  : SOP2 <op, (outs SReg_64:$dst), (ins SReg_64:$src0, SReg_64:$src1), opName, pattern>;
+
+class SOP2_VCC <bits<7> op, string opName, list<dag> pattern>
+  : SOP2 <op, (outs SReg_1:$vcc), (ins SReg_64:$src0, SReg_64:$src1), opName, pattern>;
+
+class VOP1_Helper <bits<8> op, RegisterClass vrc, RegisterClass arc,
+                   string opName, list<dag> pattern> : 
+  VOP1 <
+    op, (outs vrc:$dst), (ins arc:$src0), opName, pattern
+  >;
+
+multiclass VOP1_32 <bits<8> op, string opName, list<dag> pattern> {
+  def _e32: VOP1_Helper <op, VReg_32, AllReg_32, opName, pattern>;
+  def _e64 : VOP3_32 <{1, 1, op{6}, op{5}, op{4}, op{3}, op{2}, op{1}, op{0}},
+                      opName, []
+  >;
+}
+
+multiclass VOP1_64 <bits<8> op, string opName, list<dag> pattern> {
+
+  def _e32 : VOP1_Helper <op, VReg_64, AllReg_64, opName, pattern>;
+
+  def _e64 : VOP3_64 <
+    {1, 1, op{6}, op{5}, op{4}, op{3}, op{2}, op{1}, op{0}},
+    opName, []
+  >;
+}
+
+class VOP2_Helper <bits<6> op, RegisterClass vrc, RegisterClass arc,
+                   string opName, list<dag> pattern> :
+  VOP2 <
+    op, (outs vrc:$dst), (ins arc:$src0, vrc:$src1), opName, pattern
+  >;
+
+multiclass VOP2_32 <bits<6> op, string opName, list<dag> pattern> {
+
+  def _e32 : VOP2_Helper <op, VReg_32, AllReg_32, opName, pattern>;
+
+  def _e64 : VOP3_32 <{1, 0, 0, op{5}, op{4}, op{3}, op{2}, op{1}, op{0}},
+                      opName, []
+  >;
+}
+
+multiclass VOP2_64 <bits<6> op, string opName, list<dag> pattern> {
+  def _e32: VOP2_Helper <op, VReg_64, AllReg_64, opName, pattern>;
+
+  def _e64 : VOP3_64 <
+    {1, 0, 0, op{5}, op{4}, op{3}, op{2}, op{1}, op{0}},
+    opName, []
+  >;
+}
+
+class SOPK_32 <bits<5> op, string opName, list<dag> pattern>
+  : SOPK <op, (outs SReg_32:$dst), (ins i16imm:$src0), opName, pattern>;
+
+class SOPK_64 <bits<5> op, string opName, list<dag> pattern>
+  : SOPK <op, (outs SReg_64:$dst), (ins i16imm:$src0), opName, pattern>;
+
+class VOPC_Helper <bits<8> op, RegisterClass vrc, RegisterClass arc,
+                 string opName, list<dag> pattern> :
+  VOPC <
+    op, (ins arc:$src0, vrc:$src1), opName, pattern
+  >;
+
+multiclass VOPC_32 <bits<9> op, string opName, list<dag> pattern> {
+
+  def _e32 : VOPC_Helper <
+    {op{7}, op{6}, op{5}, op{4}, op{3}, op{2}, op{1}, op{0}},
+    VReg_32, AllReg_32, opName, pattern
+  >;
+
+  def _e64 : VOP3_1_32 <
+    op,
+    opName, pattern
+  >;
+}
+
+multiclass VOPC_64 <bits<8> op, string opName, list<dag> pattern> {
+
+  def _e32 : VOPC_Helper <op, VReg_64, AllReg_64, opName, pattern>;
+
+  def _e64 : VOP3_64 <
+    {0, op{7}, op{6}, op{5}, op{4}, op{3}, op{2}, op{1}, op{0}},
+    opName, []
+  >;
+}
+
+class SOPC_32 <bits<7> op, string opName, list<dag> pattern>
+  : SOPC <op, (outs SCCReg:$dst), (ins SReg_32:$src0, SReg_32:$src1), opName, pattern>;
+
+class SOPC_64 <bits<7> op, string opName, list<dag> pattern>
+  : SOPC <op, (outs SCCReg:$dst), (ins SReg_64:$src0, SReg_64:$src1), opName, pattern>;
+
diff --git a/lib/Target/R600/SIInstrInfo.cpp b/lib/Target/R600/SIInstrInfo.cpp
new file mode 100644
index 0000000000..4aa4a45fb5
--- /dev/null
+++ b/lib/Target/R600/SIInstrInfo.cpp
@@ -0,0 +1,89 @@
+//===-- SIInstrInfo.cpp - SI Instruction Information  ---------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief SI Implementation of TargetInstrInfo.
+//
+//===----------------------------------------------------------------------===//
+
+
+#include "SIInstrInfo.h"
+#include "AMDGPUTargetMachine.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/MC/MCInstrDesc.h"
+
+#include <stdio.h>
+
+using namespace llvm;
+
+SIInstrInfo::SIInstrInfo(AMDGPUTargetMachine &tm)
+  : AMDGPUInstrInfo(tm),
+    RI(tm, *this)
+    { }
+
+const SIRegisterInfo &SIInstrInfo::getRegisterInfo() const {
+  return RI;
+}
+
+void
+SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
+                           MachineBasicBlock::iterator MI, DebugLoc DL,
+                           unsigned DestReg, unsigned SrcReg,
+                           bool KillSrc) const {
+  // If we are trying to copy to or from SCC, there is a bug somewhere else in
+  // the backend.  While it may be theoretically possible to do this, it should
+  // never be necessary.
+  assert(DestReg != AMDGPU::SCC && SrcReg != AMDGPU::SCC);
+
+  if (AMDGPU::SReg_64RegClass.contains(DestReg)) {
+    assert(AMDGPU::SReg_64RegClass.contains(SrcReg));
+    BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
+            .addReg(SrcReg, getKillRegState(KillSrc));
+  } else if (AMDGPU::VReg_32RegClass.contains(DestReg)) {
+    assert(AMDGPU::VReg_32RegClass.contains(SrcReg) ||
+           AMDGPU::SReg_32RegClass.contains(SrcReg));
+    BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
+            .addReg(SrcReg, getKillRegState(KillSrc));
+  } else {
+    assert(AMDGPU::SReg_32RegClass.contains(DestReg));
+    assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
+    BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
+            .addReg(SrcReg, getKillRegState(KillSrc));
+  }
+}
+
+MachineInstr * SIInstrInfo::getMovImmInstr(MachineFunction *MF, unsigned DstReg,
+                                           int64_t Imm) const {
+  MachineInstr * MI = MF->CreateMachineInstr(get(AMDGPU::V_MOV_IMM_I32), DebugLoc());
+  MachineInstrBuilder(MI).addReg(DstReg, RegState::Define);
+  MachineInstrBuilder(MI).addImm(Imm);
+
+  return MI;
+
+}
+
+bool SIInstrInfo::isMov(unsigned Opcode) const {
+  switch(Opcode) {
+  default: return false;
+  case AMDGPU::S_MOV_B32:
+  case AMDGPU::S_MOV_B64:
+  case AMDGPU::V_MOV_B32_e32:
+  case AMDGPU::V_MOV_B32_e64:
+  case AMDGPU::V_MOV_IMM_F32:
+  case AMDGPU::V_MOV_IMM_I32:
+  case AMDGPU::S_MOV_IMM_I32:
+    return true;
+  }
+}
+
+bool
+SIInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
+  return RC != &AMDGPU::EXECRegRegClass;
+}
diff --git a/lib/Target/R600/SIInstrInfo.h b/lib/Target/R600/SIInstrInfo.h
new file mode 100644
index 0000000000..631f6c00cc
--- /dev/null
+++ b/lib/Target/R600/SIInstrInfo.h
@@ -0,0 +1,62 @@
+//===-- SIInstrInfo.h - SI Instruction Info Interface ---------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Interface definition for SIInstrInfo.
+//
+//===----------------------------------------------------------------------===//
+
+
+#ifndef SIINSTRINFO_H
+#define SIINSTRINFO_H
+
+#include "AMDGPUInstrInfo.h"
+#include "SIRegisterInfo.h"
+
+namespace llvm {
+
+class SIInstrInfo : public AMDGPUInstrInfo {
+private:
+  const SIRegisterInfo RI;
+
+public:
+  explicit SIInstrInfo(AMDGPUTargetMachine &tm);
+
+  const SIRegisterInfo &getRegisterInfo() const;
+
+  virtual void copyPhysReg(MachineBasicBlock &MBB,
+                           MachineBasicBlock::iterator MI, DebugLoc DL,
+                           unsigned DestReg, unsigned SrcReg,
+                           bool KillSrc) const;
+
+  /// \returns the encoding type of this instruction.
+  unsigned getEncodingType(const MachineInstr &MI) const;
+
+  /// \returns the size of this instructions encoding in number of bytes.
+  unsigned getEncodingBytes(const MachineInstr &MI) const;
+
+  virtual MachineInstr * getMovImmInstr(MachineFunction *MF, unsigned DstReg,
+                                        int64_t Imm) const;
+
+  virtual unsigned getIEQOpcode() const { assert(!"Implement"); return 0;}
+  virtual bool isMov(unsigned Opcode) const;
+
+  virtual bool isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const;
+  };
+
+} // End namespace llvm
+
+namespace SIInstrFlags {
+  enum Flags {
+    // First 4 bits are the instruction encoding
+    NEED_WAIT = 1 << 4
+  };
+}
+
+#endif //SIINSTRINFO_H
diff --git a/lib/Target/R600/SIInstrInfo.td b/lib/Target/R600/SIInstrInfo.td
new file mode 100644
index 0000000000..873a451e99
--- /dev/null
+++ b/lib/Target/R600/SIInstrInfo.td
@@ -0,0 +1,589 @@
+//===-- SIInstrInfo.td - SI Instruction Encodings ---------*- tablegen -*--===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+//===----------------------------------------------------------------------===//
+// SI DAG Profiles
+//===----------------------------------------------------------------------===//
+def SDTVCCBinaryOp : SDTypeProfile<1, 2, [
+  SDTCisInt<0>, SDTCisInt<1>, SDTCisSameAs<1, 2>
+]>;
+
+//===----------------------------------------------------------------------===//
+// SI DAG Nodes
+//===----------------------------------------------------------------------===//
+
+// and operation on 64-bit wide vcc
+def SIsreg1_and : SDNode<"SIISD::VCC_AND", SDTVCCBinaryOp,
+  [SDNPCommutative, SDNPAssociative]
+>;
+
+// Special bitcast node for sharing VCC register between VALU and SALU
+def SIsreg1_bitcast : SDNode<"SIISD::VCC_BITCAST",
+  SDTypeProfile<1, 1, [SDTCisInt<0>, SDTCisInt<1>]>
+>;
+
+// and operation on 64-bit wide vcc
+def SIvcc_and : SDNode<"SIISD::VCC_AND", SDTVCCBinaryOp,
+  [SDNPCommutative, SDNPAssociative]
+>;
+
+// Special bitcast node for sharing VCC register between VALU and SALU
+def SIvcc_bitcast : SDNode<"SIISD::VCC_BITCAST",
+  SDTypeProfile<1, 1, [SDTCisInt<0>, SDTCisInt<1>]>
+>;
+
+class InstSI <dag outs, dag ins, string asm, list<dag> pattern> :
+    AMDGPUInst<outs, ins, asm, pattern> {
+
+  field bits<4> EncodingType = 0;
+  field bits<1> NeedWait = 0;
+
+  let TSFlags{3-0} = EncodingType;
+  let TSFlags{4} = NeedWait;
+
+}
+
+class Enc32 <dag outs, dag ins, string asm, list<dag> pattern> :
+    InstSI <outs, ins, asm, pattern> {
+
+  field bits<32> Inst;
+}
+
+class Enc64 <dag outs, dag ins, string asm, list<dag> pattern> :
+    InstSI <outs, ins, asm, pattern> {
+
+  field bits<64> Inst;
+}
+
+class SIOperand <ValueType vt, dag opInfo>: Operand <vt> {
+  let EncoderMethod = "encodeOperand";
+  let MIOperandInfo = opInfo;
+}
+
+def IMM16bit : ImmLeaf <
+  i16,
+  [{return isInt<16>(Imm);}]
+>;
+
+def IMM8bit : ImmLeaf <
+  i32,
+  [{return (int32_t)Imm >= 0 && (int32_t)Imm <= 0xff;}]
+>;
+
+def IMM12bit : ImmLeaf <
+  i16,
+  [{return (int16_t)Imm >= 0 && (int16_t)Imm <= 0xfff;}]
+>;
+
+def IMM32bitIn64bit : ImmLeaf <
+  i64,
+  [{return isInt<32>(Imm);}]
+>;
+
+class GPR4Align <RegisterClass rc> : Operand <vAny> {
+  let EncoderMethod = "GPR4AlignEncode";
+  let MIOperandInfo = (ops rc:$reg); 
+}
+
+class GPR2Align <RegisterClass rc, ValueType vt> : Operand <vt> {
+  let EncoderMethod = "GPR2AlignEncode";
+  let MIOperandInfo = (ops rc:$reg);
+}
+
+def SMRDmemrr : Operand<iPTR> {
+  let MIOperandInfo = (ops SReg_64, SReg_32);
+  let EncoderMethod = "GPR2AlignEncode";
+}
+
+def SMRDmemri : Operand<iPTR> {
+  let MIOperandInfo = (ops SReg_64, i32imm);
+  let EncoderMethod = "SMRDmemriEncode";
+}
+
+def ADDR_Reg     : ComplexPattern<i64, 2, "SelectADDRReg", [], []>;
+def ADDR_Offset8 : ComplexPattern<i64, 2, "SelectADDR8BitOffset", [], []>;
+
+let Uses = [EXEC] in {
+
+def EXP : Enc64<
+  (outs),
+  (ins i32imm:$en, i32imm:$tgt, i32imm:$compr, i32imm:$done, i32imm:$vm,
+       VReg_32:$src0, VReg_32:$src1, VReg_32:$src2, VReg_32:$src3),
+  "EXP $en, $tgt, $compr, $done, $vm, $src0, $src1, $src2, $src3",
+  [] > {
+
+  bits<4> EN;
+  bits<6> TGT;
+  bits<1> COMPR;
+  bits<1> DONE;
+  bits<1> VM;
+  bits<8> VSRC0;
+  bits<8> VSRC1;
+  bits<8> VSRC2;
+  bits<8> VSRC3;
+
+  let Inst{3-0} = EN;
+  let Inst{9-4} = TGT;
+  let Inst{10} = COMPR;
+  let Inst{11} = DONE;
+  let Inst{12} = VM;
+  let Inst{31-26} = 0x3e;
+  let Inst{39-32} = VSRC0;
+  let Inst{47-40} = VSRC1;
+  let Inst{55-48} = VSRC2;
+  let Inst{63-56} = VSRC3;
+  let EncodingType = 0; //SIInstrEncodingType::EXP
+
+  let NeedWait = 1;
+  let usesCustomInserter = 1;
+}
+
+class MIMG <bits<7> op, dag outs, dag ins, string asm, list<dag> pattern> :
+    Enc64 <outs, ins, asm, pattern> {
+
+  bits<8> VDATA;
+  bits<4> DMASK;
+  bits<1> UNORM;
+  bits<1> GLC;
+  bits<1> DA;
+  bits<1> R128;
+  bits<1> TFE;
+  bits<1> LWE;
+  bits<1> SLC;
+  bits<8> VADDR;
+  bits<5> SRSRC;
+  bits<5> SSAMP; 
+
+  let Inst{11-8} = DMASK;
+  let Inst{12} = UNORM;
+  let Inst{13} = GLC;
+  let Inst{14} = DA;
+  let Inst{15} = R128;
+  let Inst{16} = TFE;
+  let Inst{17} = LWE;
+  let Inst{24-18} = op;
+  let Inst{25} = SLC;
+  let Inst{31-26} = 0x3c;
+  let Inst{39-32} = VADDR;
+  let Inst{47-40} = VDATA;
+  let Inst{52-48} = SRSRC;
+  let Inst{57-53} = SSAMP;
+
+  let EncodingType = 2; //SIInstrEncodingType::MIMG
+
+  let NeedWait = 1;
+  let usesCustomInserter = 1;
+}
+
+class MTBUF <bits<3> op, dag outs, dag ins, string asm, list<dag> pattern> :
+    Enc64<outs, ins, asm, pattern> {
+
+  bits<8> VDATA;
+  bits<12> OFFSET;
+  bits<1> OFFEN;
+  bits<1> IDXEN;
+  bits<1> GLC;
+  bits<1> ADDR64;
+  bits<4> DFMT;
+  bits<3> NFMT;
+  bits<8> VADDR;
+  bits<5> SRSRC;
+  bits<1> SLC;
+  bits<1> TFE;
+  bits<8> SOFFSET;
+
+  let Inst{11-0} = OFFSET;
+  let Inst{12} = OFFEN;
+  let Inst{13} = IDXEN;
+  let Inst{14} = GLC;
+  let Inst{15} = ADDR64;
+  let Inst{18-16} = op;
+  let Inst{22-19} = DFMT;
+  let Inst{25-23} = NFMT;
+  let Inst{31-26} = 0x3a; //encoding
+  let Inst{39-32} = VADDR;
+  let Inst{47-40} = VDATA;
+  let Inst{52-48} = SRSRC;
+  let Inst{54} = SLC;
+  let Inst{55} = TFE;
+  let Inst{63-56} = SOFFSET;
+  let EncodingType = 3; //SIInstrEncodingType::MTBUF
+
+  let NeedWait = 1;
+  let usesCustomInserter = 1;
+  let neverHasSideEffects = 1;
+}
+
+class MUBUF <bits<7> op, dag outs, dag ins, string asm, list<dag> pattern> :
+    Enc64<outs, ins, asm, pattern> {
+
+  bits<8> VDATA;
+  bits<12> OFFSET;
+  bits<1> OFFEN;
+  bits<1> IDXEN;
+  bits<1> GLC;
+  bits<1> ADDR64;
+  bits<1> LDS;
+  bits<8> VADDR;
+  bits<5> SRSRC;
+  bits<1> SLC;
+  bits<1> TFE;
+  bits<8> SOFFSET;
+
+  let Inst{11-0} = OFFSET;
+  let Inst{12} = OFFEN;
+  let Inst{13} = IDXEN;
+  let Inst{14} = GLC;
+  let Inst{15} = ADDR64;
+  let Inst{16} = LDS;
+  let Inst{24-18} = op;
+  let Inst{31-26} = 0x38; //encoding
+  let Inst{39-32} = VADDR;
+  let Inst{47-40} = VDATA;
+  let Inst{52-48} = SRSRC;
+  let Inst{54} = SLC;
+  let Inst{55} = TFE;
+  let Inst{63-56} = SOFFSET;
+  let EncodingType = 4; //SIInstrEncodingType::MUBUF
+
+  let NeedWait = 1;
+  let usesCustomInserter = 1;
+  let neverHasSideEffects = 1;
+}
+
+} // End Uses = [EXEC]
+
+class SMRD <bits<5> op, dag outs, dag ins, string asm, list<dag> pattern> :
+    Enc32<outs, ins, asm, pattern> {
+
+  bits<7> SDST;
+  bits<15> PTR;
+  bits<8> OFFSET = PTR{7-0};
+  bits<1> IMM    = PTR{8};
+  bits<6> SBASE  = PTR{14-9};
+  
+  let Inst{7-0} = OFFSET;
+  let Inst{8} = IMM;
+  let Inst{14-9} = SBASE;
+  let Inst{21-15} = SDST;
+  let Inst{26-22} = op;
+  let Inst{31-27} = 0x18; //encoding
+  let EncodingType = 5; //SIInstrEncodingType::SMRD
+
+  let NeedWait = 1;
+  let usesCustomInserter = 1;
+}
+
+class SOP1 <bits<8> op, dag outs, dag ins, string asm, list<dag> pattern> :
+    Enc32<outs, ins, asm, pattern> {
+
+  bits<7> SDST;
+  bits<8> SSRC0;
+
+  let Inst{7-0} = SSRC0;
+  let Inst{15-8} = op;
+  let Inst{22-16} = SDST;
+  let Inst{31-23} = 0x17d; //encoding;
+  let EncodingType = 6; //SIInstrEncodingType::SOP1
+
+  let mayLoad = 0;
+  let mayStore = 0;
+  let hasSideEffects = 0;
+}
+
+class SOP2 <bits<7> op, dag outs, dag ins, string asm, list<dag> pattern> :
+    Enc32 <outs, ins, asm, pattern> {
+  
+  bits<7> SDST;
+  bits<8> SSRC0;
+  bits<8> SSRC1;
+
+  let Inst{7-0} = SSRC0;
+  let Inst{15-8} = SSRC1;
+  let Inst{22-16} = SDST;
+  let Inst{29-23} = op;
+  let Inst{31-30} = 0x2; // encoding
+  let EncodingType = 7; // SIInstrEncodingType::SOP2  
+
+  let mayLoad = 0;
+  let mayStore = 0;
+  let hasSideEffects = 0;
+}
+
+class SOPC <bits<7> op, dag outs, dag ins, string asm, list<dag> pattern> :
+  Enc32<outs, ins, asm, pattern> {
+
+  bits<8> SSRC0;
+  bits<8> SSRC1;
+
+  let Inst{7-0} = SSRC0;
+  let Inst{15-8} = SSRC1;
+  let Inst{22-16} = op;
+  let Inst{31-23} = 0x17e;
+  let EncodingType = 8; // SIInstrEncodingType::SOPC
+
+  let DisableEncoding = "$dst";
+  let mayLoad = 0;
+  let mayStore = 0;
+  let hasSideEffects = 0;
+}
+
+class SOPK <bits<5> op, dag outs, dag ins, string asm, list<dag> pattern> :
+   Enc32 <outs, ins , asm, pattern> {
+
+  bits <7> SDST;
+  bits <16> SIMM16;
+  
+  let Inst{15-0} = SIMM16;
+  let Inst{22-16} = SDST;
+  let Inst{27-23} = op;
+  let Inst{31-28} = 0xb; //encoding
+  let EncodingType = 9; // SIInstrEncodingType::SOPK
+
+  let mayLoad = 0;
+  let mayStore = 0;
+  let hasSideEffects = 0;
+}
+
+class SOPP <bits<7> op, dag ins, string asm, list<dag> pattern> : Enc32 <
+  (outs),
+  ins,
+  asm,
+  pattern > {
+
+  bits <16> SIMM16;
+
+  let Inst{15-0} = SIMM16;
+  let Inst{22-16} = op;
+  let Inst{31-23} = 0x17f; // encoding
+  let EncodingType = 10; // SIInstrEncodingType::SOPP
+
+  let mayLoad = 0;
+  let mayStore = 0;
+  let hasSideEffects = 0;
+}
+    
+let Uses = [EXEC] in {
+
+class VINTRP <bits <2> op, dag outs, dag ins, string asm, list<dag> pattern> :
+    Enc32 <outs, ins, asm, pattern> {
+
+  bits<8> VDST;
+  bits<8> VSRC;
+  bits<2> ATTRCHAN;
+  bits<6> ATTR;
+
+  let Inst{7-0} = VSRC;
+  let Inst{9-8} = ATTRCHAN;
+  let Inst{15-10} = ATTR;
+  let Inst{17-16} = op;
+  let Inst{25-18} = VDST;
+  let Inst{31-26} = 0x32; // encoding
+  let EncodingType = 11; // SIInstrEncodingType::VINTRP
+
+  let neverHasSideEffects = 1;
+  let mayLoad = 1;
+  let mayStore = 0;
+}
+
+class VOP1 <bits<8> op, dag outs, dag ins, string asm, list<dag> pattern> :
+    Enc32 <outs, ins, asm, pattern> {
+
+  bits<8> VDST;
+  bits<9> SRC0;
+  
+  let Inst{8-0} = SRC0;
+  let Inst{16-9} = op;
+  let Inst{24-17} = VDST;
+  let Inst{31-25} = 0x3f; //encoding
+  
+  let EncodingType = 12; // SIInstrEncodingType::VOP1
+  let PostEncoderMethod = "VOPPostEncode";
+
+  let mayLoad = 0;
+  let mayStore = 0;
+  let hasSideEffects = 0;
+}
+
+class VOP2 <bits<6> op, dag outs, dag ins, string asm, list<dag> pattern> :
+    Enc32 <outs, ins, asm, pattern> {
+
+  bits<8> VDST;
+  bits<9> SRC0;
+  bits<8> VSRC1;
+  
+  let Inst{8-0} = SRC0;
+  let Inst{16-9} = VSRC1;
+  let Inst{24-17} = VDST;
+  let Inst{30-25} = op;
+  let Inst{31} = 0x0; //encoding
+  
+  let EncodingType = 13; // SIInstrEncodingType::VOP2
+  let PostEncoderMethod = "VOPPostEncode";
+
+  let mayLoad = 0;
+  let mayStore = 0;
+  let hasSideEffects = 0;
+}
+
+class VOP3 <bits<9> op, dag outs, dag ins, string asm, list<dag> pattern> :
+    Enc64 <outs, ins, asm, pattern> {
+
+  bits<8> VDST;
+  bits<9> SRC0;
+  bits<9> SRC1;
+  bits<9> SRC2;
+  bits<3> ABS; 
+  bits<1> CLAMP;
+  bits<2> OMOD;
+  bits<3> NEG;
+
+  let Inst{7-0} = VDST;
+  let Inst{10-8} = ABS;
+  let Inst{11} = CLAMP;
+  let Inst{25-17} = op;
+  let Inst{31-26} = 0x34; //encoding
+  let Inst{40-32} = SRC0;
+  let Inst{49-41} = SRC1;
+  let Inst{58-50} = SRC2;
+  let Inst{60-59} = OMOD;
+  let Inst{63-61} = NEG;
+  
+  let EncodingType = 14; // SIInstrEncodingType::VOP3
+  let PostEncoderMethod = "VOPPostEncode";
+
+  let mayLoad = 0;
+  let mayStore = 0;
+  let hasSideEffects = 0;
+}
+
+class VOP3b <bits<9> op, dag outs, dag ins, string asm, list<dag> pattern> :
+    Enc64 <outs, ins, asm, pattern> {
+
+  bits<8> VDST;
+  bits<9> SRC0;
+  bits<9> SRC1;
+  bits<9> SRC2;
+  bits<7> SDST;
+  bits<2> OMOD;
+  bits<3> NEG;
+
+  let Inst{7-0} = VDST;
+  let Inst{14-8} = SDST;
+  let Inst{25-17} = op;
+  let Inst{31-26} = 0x34; //encoding
+  let Inst{40-32} = SRC0;
+  let Inst{49-41} = SRC1;
+  let Inst{58-50} = SRC2;
+  let Inst{60-59} = OMOD;
+  let Inst{63-61} = NEG;
+
+  let EncodingType = 14; // SIInstrEncodingType::VOP3
+  let PostEncoderMethod = "VOPPostEncode";
+
+  let mayLoad = 0;
+  let mayStore = 0;
+  let hasSideEffects = 0;
+}
+
+class VOPC <bits<8> op, dag ins, string asm, list<dag> pattern> :
+    Enc32 <(outs VCCReg:$dst), ins, asm, pattern> {
+
+  bits<9> SRC0;
+  bits<8> VSRC1;
+
+  let Inst{8-0} = SRC0;
+  let Inst{16-9} = VSRC1;
+  let Inst{24-17} = op;
+  let Inst{31-25} = 0x3e;
+ 
+  let EncodingType = 15; //SIInstrEncodingType::VOPC
+  let PostEncoderMethod = "VOPPostEncode";
+  let DisableEncoding = "$dst";
+  let mayLoad = 0;
+  let mayStore = 0;
+  let hasSideEffects = 0;
+}
+
+} // End Uses = [EXEC]
+
+class MIMG_Load_Helper <bits<7> op, string asm> : MIMG <
+  op,
+  (outs VReg_128:$vdata),
+  (ins i32imm:$dmask, i1imm:$unorm, i1imm:$glc, i1imm:$da, i1imm:$r128,
+       i1imm:$tfe, i1imm:$lwe, i1imm:$slc, VReg_128:$vaddr,
+       GPR4Align<SReg_256>:$srsrc, GPR4Align<SReg_128>:$ssamp),
+  asm,
+  []> {
+  let mayLoad = 1;
+  let mayStore = 0;
+}
+
+class MUBUF_Load_Helper <bits<7> op, string asm, RegisterClass regClass> : MUBUF <
+  op,
+  (outs regClass:$dst),
+  (ins i16imm:$offset, i1imm:$offen, i1imm:$idxen, i1imm:$glc, i1imm:$addr64,
+       i1imm:$lds, VReg_32:$vaddr, GPR4Align<SReg_128>:$srsrc, i1imm:$slc,
+       i1imm:$tfe, SReg_32:$soffset),
+  asm,
+  []> {
+  let mayLoad = 1;
+  let mayStore = 0;
+}
+
+class MTBUF_Load_Helper <bits<3> op, string asm, RegisterClass regClass> : MTBUF <
+  op,
+  (outs regClass:$dst),
+  (ins i16imm:$offset, i1imm:$offen, i1imm:$idxen, i1imm:$glc, i1imm:$addr64,
+       i8imm:$dfmt, i8imm:$nfmt, VReg_32:$vaddr, GPR4Align<SReg_128>:$srsrc,
+       i1imm:$slc, i1imm:$tfe, SReg_32:$soffset),
+  asm,
+  []> {
+  let mayLoad = 1;
+  let mayStore = 0;
+}
+
+class MTBUF_Store_Helper <bits<3> op, string asm, RegisterClass regClass> : MTBUF <
+  op,
+  (outs),
+  (ins regClass:$vdata, i16imm:$offset, i1imm:$offen, i1imm:$idxen, i1imm:$glc,
+   i1imm:$addr64, i8imm:$dfmt, i8imm:$nfmt, VReg_32:$vaddr,
+   GPR4Align<SReg_128>:$srsrc, i1imm:$slc, i1imm:$tfe, SReg_32:$soffset),
+  asm,
+  []> {
+  let mayStore = 1;
+  let mayLoad = 0;
+}
+
+multiclass SMRD_Helper <bits<5> op, string asm, RegisterClass dstClass,
+                        ValueType vt> {
+  def _IMM : SMRD <
+              op,
+              (outs dstClass:$dst),
+              (ins SMRDmemri:$src0),
+              asm,
+              [(set (vt dstClass:$dst), (constant_load ADDR_Offset8:$src0))]
+  >;
+
+  def _SGPR : SMRD <
+              op,
+              (outs dstClass:$dst),
+              (ins SMRDmemrr:$src0),
+              asm,
+              [(set (vt dstClass:$dst), (constant_load ADDR_Reg:$src0))]
+  >;
+}
+
+multiclass SMRD_32 <bits<5> op, string asm, RegisterClass dstClass> {
+  defm _F32 : SMRD_Helper <op, asm, dstClass, f32>;
+  defm _I32 : SMRD_Helper <op, asm, dstClass, i32>;
+}
+
+include "SIInstrFormats.td"
+include "SIInstructions.td"
diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
new file mode 100644
index 0000000000..ea8de91406
--- /dev/null
+++ b/lib/Target/R600/SIInstructions.td
@@ -0,0 +1,1306 @@
+//===-- SIInstructions.td - SI Instruction Defintions ---------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+// This file was originally auto-generated from a GPU register header file and
+// all the instruction definitions were originally commented out.  Instructions
+// that are not yet supported remain commented out.
+//===----------------------------------------------------------------------===//
+
+def isSI : Predicate<"Subtarget.device()"
+                            "->getGeneration() == AMDGPUDeviceInfo::HD7XXX">;
+
+let Predicates = [isSI] in {
+
+let neverHasSideEffects = 1 in {
+def S_MOV_B32 : SOP1_32 <0x00000003, "S_MOV_B32", []>;
+def S_MOV_B64 : SOP1_64 <0x00000004, "S_MOV_B64", []>;
+def S_CMOV_B32 : SOP1_32 <0x00000005, "S_CMOV_B32", []>;
+def S_CMOV_B64 : SOP1_64 <0x00000006, "S_CMOV_B64", []>;
+def S_NOT_B32 : SOP1_32 <0x00000007, "S_NOT_B32", []>;
+def S_NOT_B64 : SOP1_64 <0x00000008, "S_NOT_B64", []>;
+def S_WQM_B32 : SOP1_32 <0x00000009, "S_WQM_B32", []>;
+def S_WQM_B64 : SOP1_64 <0x0000000a, "S_WQM_B64", []>;
+def S_BREV_B32 : SOP1_32 <0x0000000b, "S_BREV_B32", []>;
+def S_BREV_B64 : SOP1_64 <0x0000000c, "S_BREV_B64", []>;
+} // End neverHasSideEffects = 1
+////def S_BCNT0_I32_B32 : SOP1_BCNT0 <0x0000000d, "S_BCNT0_I32_B32", []>;
+////def S_BCNT0_I32_B64 : SOP1_BCNT0 <0x0000000e, "S_BCNT0_I32_B64", []>;
+////def S_BCNT1_I32_B32 : SOP1_BCNT1 <0x0000000f, "S_BCNT1_I32_B32", []>;
+////def S_BCNT1_I32_B64 : SOP1_BCNT1 <0x00000010, "S_BCNT1_I32_B64", []>;
+////def S_FF0_I32_B32 : SOP1_FF0 <0x00000011, "S_FF0_I32_B32", []>;
+////def S_FF0_I32_B64 : SOP1_FF0 <0x00000012, "S_FF0_I32_B64", []>;
+////def S_FF1_I32_B32 : SOP1_FF1 <0x00000013, "S_FF1_I32_B32", []>;
+////def S_FF1_I32_B64 : SOP1_FF1 <0x00000014, "S_FF1_I32_B64", []>;
+//def S_FLBIT_I32_B32 : SOP1_32 <0x00000015, "S_FLBIT_I32_B32", []>;
+//def S_FLBIT_I32_B64 : SOP1_32 <0x00000016, "S_FLBIT_I32_B64", []>;
+def S_FLBIT_I32 : SOP1_32 <0x00000017, "S_FLBIT_I32", []>;
+//def S_FLBIT_I32_I64 : SOP1_32 <0x00000018, "S_FLBIT_I32_I64", []>;
+//def S_SEXT_I32_I8 : SOP1_32 <0x00000019, "S_SEXT_I32_I8", []>;
+//def S_SEXT_I32_I16 : SOP1_32 <0x0000001a, "S_SEXT_I32_I16", []>;
+////def S_BITSET0_B32 : SOP1_BITSET0 <0x0000001b, "S_BITSET0_B32", []>;
+////def S_BITSET0_B64 : SOP1_BITSET0 <0x0000001c, "S_BITSET0_B64", []>;
+////def S_BITSET1_B32 : SOP1_BITSET1 <0x0000001d, "S_BITSET1_B32", []>;
+////def S_BITSET1_B64 : SOP1_BITSET1 <0x0000001e, "S_BITSET1_B64", []>;
+def S_GETPC_B64 : SOP1_64 <0x0000001f, "S_GETPC_B64", []>;
+def S_SETPC_B64 : SOP1_64 <0x00000020, "S_SETPC_B64", []>;
+def S_SWAPPC_B64 : SOP1_64 <0x00000021, "S_SWAPPC_B64", []>;
+def S_RFE_B64 : SOP1_64 <0x00000022, "S_RFE_B64", []>;
+
+let hasSideEffects = 1, Uses = [EXEC], Defs = [EXEC] in {
+
+def S_AND_SAVEEXEC_B64 : SOP1_64 <0x00000024, "S_AND_SAVEEXEC_B64", []>;
+def S_OR_SAVEEXEC_B64 : SOP1_64 <0x00000025, "S_OR_SAVEEXEC_B64", []>;
+def S_XOR_SAVEEXEC_B64 : SOP1_64 <0x00000026, "S_XOR_SAVEEXEC_B64", []>;
+def S_ANDN2_SAVEEXEC_B64 : SOP1_64 <0x00000027, "S_ANDN2_SAVEEXEC_B64", []>;
+def S_ORN2_SAVEEXEC_B64 : SOP1_64 <0x00000028, "S_ORN2_SAVEEXEC_B64", []>;
+def S_NAND_SAVEEXEC_B64 : SOP1_64 <0x00000029, "S_NAND_SAVEEXEC_B64", []>;
+def S_NOR_SAVEEXEC_B64 : SOP1_64 <0x0000002a, "S_NOR_SAVEEXEC_B64", []>;
+def S_XNOR_SAVEEXEC_B64 : SOP1_64 <0x0000002b, "S_XNOR_SAVEEXEC_B64", []>;
+
+} // End hasSideEffects = 1
+
+def S_QUADMASK_B32 : SOP1_32 <0x0000002c, "S_QUADMASK_B32", []>;
+def S_QUADMASK_B64 : SOP1_64 <0x0000002d, "S_QUADMASK_B64", []>;
+def S_MOVRELS_B32 : SOP1_32 <0x0000002e, "S_MOVRELS_B32", []>;
+def S_MOVRELS_B64 : SOP1_64 <0x0000002f, "S_MOVRELS_B64", []>;
+def S_MOVRELD_B32 : SOP1_32 <0x00000030, "S_MOVRELD_B32", []>;
+def S_MOVRELD_B64 : SOP1_64 <0x00000031, "S_MOVRELD_B64", []>;
+//def S_CBRANCH_JOIN : SOP1_ <0x00000032, "S_CBRANCH_JOIN", []>;
+def S_MOV_REGRD_B32 : SOP1_32 <0x00000033, "S_MOV_REGRD_B32", []>;
+def S_ABS_I32 : SOP1_32 <0x00000034, "S_ABS_I32", []>;
+def S_MOV_FED_B32 : SOP1_32 <0x00000035, "S_MOV_FED_B32", []>;
+def S_MOVK_I32 : SOPK_32 <0x00000000, "S_MOVK_I32", []>;
+def S_CMOVK_I32 : SOPK_32 <0x00000002, "S_CMOVK_I32", []>;
+
+/*
+This instruction is disabled for now until we can figure out how to teach
+the instruction selector to correctly use the  S_CMP* vs V_CMP*
+instructions.
+
+When this instruction is enabled the code generator sometimes produces this
+invalid sequence:
+
+SCC = S_CMPK_EQ_I32 SGPR0, imm
+VCC = COPY SCC
+VGPR0 = V_CNDMASK VCC, VGPR0, VGPR1
+
+def S_CMPK_EQ_I32 : SOPK <
+  0x00000003, (outs SCCReg:$dst), (ins SReg_32:$src0, i32imm:$src1),
+  "S_CMPK_EQ_I32",
+  [(set SCCReg:$dst, (setcc SReg_32:$src0, imm:$src1, SETEQ))]
+>;
+*/
+
+def S_CMPK_LG_I32 : SOPK_32 <0x00000004, "S_CMPK_LG_I32", []>;
+def S_CMPK_GT_I32 : SOPK_32 <0x00000005, "S_CMPK_GT_I32", []>;
+def S_CMPK_GE_I32 : SOPK_32 <0x00000006, "S_CMPK_GE_I32", []>;
+def S_CMPK_LT_I32 : SOPK_32 <0x00000007, "S_CMPK_LT_I32", []>;
+def S_CMPK_LE_I32 : SOPK_32 <0x00000008, "S_CMPK_LE_I32", []>;
+def S_CMPK_EQ_U32 : SOPK_32 <0x00000009, "S_CMPK_EQ_U32", []>;
+def S_CMPK_LG_U32 : SOPK_32 <0x0000000a, "S_CMPK_LG_U32", []>;
+def S_CMPK_GT_U32 : SOPK_32 <0x0000000b, "S_CMPK_GT_U32", []>;
+def S_CMPK_GE_U32 : SOPK_32 <0x0000000c, "S_CMPK_GE_U32", []>;
+def S_CMPK_LT_U32 : SOPK_32 <0x0000000d, "S_CMPK_LT_U32", []>;
+def S_CMPK_LE_U32 : SOPK_32 <0x0000000e, "S_CMPK_LE_U32", []>;
+def S_ADDK_I32 : SOPK_32 <0x0000000f, "S_ADDK_I32", []>;
+def S_MULK_I32 : SOPK_32 <0x00000010, "S_MULK_I32", []>;
+//def S_CBRANCH_I_FORK : SOPK_ <0x00000011, "S_CBRANCH_I_FORK", []>;
+def S_GETREG_B32 : SOPK_32 <0x00000012, "S_GETREG_B32", []>;
+def S_SETREG_B32 : SOPK_32 <0x00000013, "S_SETREG_B32", []>;
+def S_GETREG_REGRD_B32 : SOPK_32 <0x00000014, "S_GETREG_REGRD_B32", []>;
+//def S_SETREG_IMM32_B32 : SOPK_32 <0x00000015, "S_SETREG_IMM32_B32", []>;
+//def EXP : EXP_ <0x00000000, "EXP", []>;
+
+defm V_CMP_F_F32 : VOPC_32 <0x00000000, "V_CMP_F_F32", []>;
+defm V_CMP_LT_F32 : VOPC_32 <0x00000001, "V_CMP_LT_F32", []>;
+def : Pat <
+  (i1 (setcc (f32 AllReg_32:$src0), VReg_32:$src1, COND_LT)),
+  (V_CMP_LT_F32_e64 AllReg_32:$src0, VReg_32:$src1)
+>;
+defm V_CMP_EQ_F32 : VOPC_32 <0x00000002, "V_CMP_EQ_F32", []>;
+def : Pat <
+  (i1 (setcc (f32 AllReg_32:$src0), VReg_32:$src1, COND_EQ)),
+  (V_CMP_EQ_F32_e64 AllReg_32:$src0, VReg_32:$src1)
+>;
+defm V_CMP_LE_F32 : VOPC_32 <0x00000003, "V_CMP_LE_F32", []>;
+def : Pat <
+  (i1 (setcc (f32 AllReg_32:$src0), VReg_32:$src1, COND_LE)),
+  (V_CMP_LE_F32_e64 AllReg_32:$src0, VReg_32:$src1)
+>;
+defm V_CMP_GT_F32 : VOPC_32 <0x00000004, "V_CMP_GT_F32", []>;
+def : Pat <
+  (i1 (setcc (f32 AllReg_32:$src0), VReg_32:$src1, COND_GT)),
+  (V_CMP_GT_F32_e64 AllReg_32:$src0, VReg_32:$src1)
+>;
+defm V_CMP_LG_F32 : VOPC_32 <0x00000005, "V_CMP_LG_F32", []>;
+def : Pat <
+  (i1 (setcc (f32 AllReg_32:$src0), VReg_32:$src1, COND_NE)),
+  (V_CMP_LG_F32_e64 AllReg_32:$src0, VReg_32:$src1)
+>;
+defm V_CMP_GE_F32 : VOPC_32 <0x00000006, "V_CMP_GE_F32", []>;
+def : Pat <
+  (i1 (setcc (f32 AllReg_32:$src0), VReg_32:$src1, COND_GE)),
+  (V_CMP_GE_F32_e64 AllReg_32:$src0, VReg_32:$src1)
+>;
+defm V_CMP_O_F32 : VOPC_32 <0x00000007, "V_CMP_O_F32", []>;
+defm V_CMP_U_F32 : VOPC_32 <0x00000008, "V_CMP_U_F32", []>;
+defm V_CMP_NGE_F32 : VOPC_32 <0x00000009, "V_CMP_NGE_F32", []>;
+defm V_CMP_NLG_F32 : VOPC_32 <0x0000000a, "V_CMP_NLG_F32", []>;
+defm V_CMP_NGT_F32 : VOPC_32 <0x0000000b, "V_CMP_NGT_F32", []>;
+defm V_CMP_NLE_F32 : VOPC_32 <0x0000000c, "V_CMP_NLE_F32", []>;
+defm V_CMP_NEQ_F32 : VOPC_32 <0x0000000d, "V_CMP_NEQ_F32", []>;
+def : Pat <
+  (i1 (setcc (f32 AllReg_32:$src0), VReg_32:$src1, COND_NE)),
+  (V_CMP_NEQ_F32_e64 AllReg_32:$src0, VReg_32:$src1)
+>;
+defm V_CMP_NLT_F32 : VOPC_32 <0x0000000e, "V_CMP_NLT_F32", []>;
+defm V_CMP_TRU_F32 : VOPC_32 <0x0000000f, "V_CMP_TRU_F32", []>;
+
+//Side effect is writing to EXEC
+let hasSideEffects = 1 in {
+
+defm V_CMPX_F_F32 : VOPC_32 <0x00000010, "V_CMPX_F_F32", []>;
+defm V_CMPX_LT_F32 : VOPC_32 <0x00000011, "V_CMPX_LT_F32", []>;
+defm V_CMPX_EQ_F32 : VOPC_32 <0x00000012, "V_CMPX_EQ_F32", []>;
+defm V_CMPX_LE_F32 : VOPC_32 <0x00000013, "V_CMPX_LE_F32", []>;
+defm V_CMPX_GT_F32 : VOPC_32 <0x00000014, "V_CMPX_GT_F32", []>;
+defm V_CMPX_LG_F32 : VOPC_32 <0x00000015, "V_CMPX_LG_F32", []>;
+defm V_CMPX_GE_F32 : VOPC_32 <0x00000016, "V_CMPX_GE_F32", []>;
+defm V_CMPX_O_F32 : VOPC_32 <0x00000017, "V_CMPX_O_F32", []>;
+defm V_CMPX_U_F32 : VOPC_32 <0x00000018, "V_CMPX_U_F32", []>;
+defm V_CMPX_NGE_F32 : VOPC_32 <0x00000019, "V_CMPX_NGE_F32", []>;
+defm V_CMPX_NLG_F32 : VOPC_32 <0x0000001a, "V_CMPX_NLG_F32", []>;
+defm V_CMPX_NGT_F32 : VOPC_32 <0x0000001b, "V_CMPX_NGT_F32", []>;
+defm V_CMPX_NLE_F32 : VOPC_32 <0x0000001c, "V_CMPX_NLE_F32", []>;
+defm V_CMPX_NEQ_F32 : VOPC_32 <0x0000001d, "V_CMPX_NEQ_F32", []>;
+defm V_CMPX_NLT_F32 : VOPC_32 <0x0000001e, "V_CMPX_NLT_F32", []>;
+defm V_CMPX_TRU_F32 : VOPC_32 <0x0000001f, "V_CMPX_TRU_F32", []>;
+
+} // End hasSideEffects = 1
+
+defm V_CMP_F_F64 : VOPC_64 <0x00000020, "V_CMP_F_F64", []>;
+defm V_CMP_LT_F64 : VOPC_64 <0x00000021, "V_CMP_LT_F64", []>;
+defm V_CMP_EQ_F64 : VOPC_64 <0x00000022, "V_CMP_EQ_F64", []>;
+defm V_CMP_LE_F64 : VOPC_64 <0x00000023, "V_CMP_LE_F64", []>;
+defm V_CMP_GT_F64 : VOPC_64 <0x00000024, "V_CMP_GT_F64", []>;
+defm V_CMP_LG_F64 : VOPC_64 <0x00000025, "V_CMP_LG_F64", []>;
+defm V_CMP_GE_F64 : VOPC_64 <0x00000026, "V_CMP_GE_F64", []>;
+defm V_CMP_O_F64 : VOPC_64 <0x00000027, "V_CMP_O_F64", []>;
+defm V_CMP_U_F64 : VOPC_64 <0x00000028, "V_CMP_U_F64", []>;
+defm V_CMP_NGE_F64 : VOPC_64 <0x00000029, "V_CMP_NGE_F64", []>;
+defm V_CMP_NLG_F64 : VOPC_64 <0x0000002a, "V_CMP_NLG_F64", []>;
+defm V_CMP_NGT_F64 : VOPC_64 <0x0000002b, "V_CMP_NGT_F64", []>;
+defm V_CMP_NLE_F64 : VOPC_64 <0x0000002c, "V_CMP_NLE_F64", []>;
+defm V_CMP_NEQ_F64 : VOPC_64 <0x0000002d, "V_CMP_NEQ_F64", []>;
+defm V_CMP_NLT_F64 : VOPC_64 <0x0000002e, "V_CMP_NLT_F64", []>;
+defm V_CMP_TRU_F64 : VOPC_64 <0x0000002f, "V_CMP_TRU_F64", []>;
+
+//Side effect is writing to EXEC
+let hasSideEffects = 1 in {
+
+defm V_CMPX_F_F64 : VOPC_64 <0x00000030, "V_CMPX_F_F64", []>;
+defm V_CMPX_LT_F64 : VOPC_64 <0x00000031, "V_CMPX_LT_F64", []>;
+defm V_CMPX_EQ_F64 : VOPC_64 <0x00000032, "V_CMPX_EQ_F64", []>;
+defm V_CMPX_LE_F64 : VOPC_64 <0x00000033, "V_CMPX_LE_F64", []>;
+defm V_CMPX_GT_F64 : VOPC_64 <0x00000034, "V_CMPX_GT_F64", []>;
+defm V_CMPX_LG_F64 : VOPC_64 <0x00000035, "V_CMPX_LG_F64", []>;
+defm V_CMPX_GE_F64 : VOPC_64 <0x00000036, "V_CMPX_GE_F64", []>;
+defm V_CMPX_O_F64 : VOPC_64 <0x00000037, "V_CMPX_O_F64", []>;
+defm V_CMPX_U_F64 : VOPC_64 <0x00000038, "V_CMPX_U_F64", []>;
+defm V_CMPX_NGE_F64 : VOPC_64 <0x00000039, "V_CMPX_NGE_F64", []>;
+defm V_CMPX_NLG_F64 : VOPC_64 <0x0000003a, "V_CMPX_NLG_F64", []>;
+defm V_CMPX_NGT_F64 : VOPC_64 <0x0000003b, "V_CMPX_NGT_F64", []>;
+defm V_CMPX_NLE_F64 : VOPC_64 <0x0000003c, "V_CMPX_NLE_F64", []>;
+defm V_CMPX_NEQ_F64 : VOPC_64 <0x0000003d, "V_CMPX_NEQ_F64", []>;
+defm V_CMPX_NLT_F64 : VOPC_64 <0x0000003e, "V_CMPX_NLT_F64", []>;
+defm V_CMPX_TRU_F64 : VOPC_64 <0x0000003f, "V_CMPX_TRU_F64", []>;
+
+} // End hasSideEffects = 1
+
+defm V_CMPS_F_F32 : VOPC_32 <0x00000040, "V_CMPS_F_F32", []>;
+defm V_CMPS_LT_F32 : VOPC_32 <0x00000041, "V_CMPS_LT_F32", []>;
+defm V_CMPS_EQ_F32 : VOPC_32 <0x00000042, "V_CMPS_EQ_F32", []>;
+defm V_CMPS_LE_F32 : VOPC_32 <0x00000043, "V_CMPS_LE_F32", []>;
+defm V_CMPS_GT_F32 : VOPC_32 <0x00000044, "V_CMPS_GT_F32", []>;
+defm V_CMPS_LG_F32 : VOPC_32 <0x00000045, "V_CMPS_LG_F32", []>;
+defm V_CMPS_GE_F32 : VOPC_32 <0x00000046, "V_CMPS_GE_F32", []>;
+defm V_CMPS_O_F32 : VOPC_32 <0x00000047, "V_CMPS_O_F32", []>;
+defm V_CMPS_U_F32 : VOPC_32 <0x00000048, "V_CMPS_U_F32", []>;
+defm V_CMPS_NGE_F32 : VOPC_32 <0x00000049, "V_CMPS_NGE_F32", []>;
+defm V_CMPS_NLG_F32 : VOPC_32 <0x0000004a, "V_CMPS_NLG_F32", []>;
+defm V_CMPS_NGT_F32 : VOPC_32 <0x0000004b, "V_CMPS_NGT_F32", []>;
+defm V_CMPS_NLE_F32 : VOPC_32 <0x0000004c, "V_CMPS_NLE_F32", []>;
+defm V_CMPS_NEQ_F32 : VOPC_32 <0x0000004d, "V_CMPS_NEQ_F32", []>;
+defm V_CMPS_NLT_F32 : VOPC_32 <0x0000004e, "V_CMPS_NLT_F32", []>;
+defm V_CMPS_TRU_F32 : VOPC_32 <0x0000004f, "V_CMPS_TRU_F32", []>;
+defm V_CMPSX_F_F32 : VOPC_32 <0x00000050, "V_CMPSX_F_F32", []>;
+defm V_CMPSX_LT_F32 : VOPC_32 <0x00000051, "V_CMPSX_LT_F32", []>;
+defm V_CMPSX_EQ_F32 : VOPC_32 <0x00000052, "V_CMPSX_EQ_F32", []>;
+defm V_CMPSX_LE_F32 : VOPC_32 <0x00000053, "V_CMPSX_LE_F32", []>;
+defm V_CMPSX_GT_F32 : VOPC_32 <0x00000054, "V_CMPSX_GT_F32", []>;
+defm V_CMPSX_LG_F32 : VOPC_32 <0x00000055, "V_CMPSX_LG_F32", []>;
+defm V_CMPSX_GE_F32 : VOPC_32 <0x00000056, "V_CMPSX_GE_F32", []>;
+defm V_CMPSX_O_F32 : VOPC_32 <0x00000057, "V_CMPSX_O_F32", []>;
+defm V_CMPSX_U_F32 : VOPC_32 <0x00000058, "V_CMPSX_U_F32", []>;
+defm V_CMPSX_NGE_F32 : VOPC_32 <0x00000059, "V_CMPSX_NGE_F32", []>;
+defm V_CMPSX_NLG_F32 : VOPC_32 <0x0000005a, "V_CMPSX_NLG_F32", []>;
+defm V_CMPSX_NGT_F32 : VOPC_32 <0x0000005b, "V_CMPSX_NGT_F32", []>;
+defm V_CMPSX_NLE_F32 : VOPC_32 <0x0000005c, "V_CMPSX_NLE_F32", []>;
+defm V_CMPSX_NEQ_F32 : VOPC_32 <0x0000005d, "V_CMPSX_NEQ_F32", []>;
+defm V_CMPSX_NLT_F32 : VOPC_32 <0x0000005e, "V_CMPSX_NLT_F32", []>;
+defm V_CMPSX_TRU_F32 : VOPC_32 <0x0000005f, "V_CMPSX_TRU_F32", []>;
+defm V_CMPS_F_F64 : VOPC_64 <0x00000060, "V_CMPS_F_F64", []>;
+defm V_CMPS_LT_F64 : VOPC_64 <0x00000061, "V_CMPS_LT_F64", []>;
+defm V_CMPS_EQ_F64 : VOPC_64 <0x00000062, "V_CMPS_EQ_F64", []>;
+defm V_CMPS_LE_F64 : VOPC_64 <0x00000063, "V_CMPS_LE_F64", []>;
+defm V_CMPS_GT_F64 : VOPC_64 <0x00000064, "V_CMPS_GT_F64", []>;
+defm V_CMPS_LG_F64 : VOPC_64 <0x00000065, "V_CMPS_LG_F64", []>;
+defm V_CMPS_GE_F64 : VOPC_64 <0x00000066, "V_CMPS_GE_F64", []>;
+defm V_CMPS_O_F64 : VOPC_64 <0x00000067, "V_CMPS_O_F64", []>;
+defm V_CMPS_U_F64 : VOPC_64 <0x00000068, "V_CMPS_U_F64", []>;
+defm V_CMPS_NGE_F64 : VOPC_64 <0x00000069, "V_CMPS_NGE_F64", []>;
+defm V_CMPS_NLG_F64 : VOPC_64 <0x0000006a, "V_CMPS_NLG_F64", []>;
+defm V_CMPS_NGT_F64 : VOPC_64 <0x0000006b, "V_CMPS_NGT_F64", []>;
+defm V_CMPS_NLE_F64 : VOPC_64 <0x0000006c, "V_CMPS_NLE_F64", []>;
+defm V_CMPS_NEQ_F64 : VOPC_64 <0x0000006d, "V_CMPS_NEQ_F64", []>;
+defm V_CMPS_NLT_F64 : VOPC_64 <0x0000006e, "V_CMPS_NLT_F64", []>;
+defm V_CMPS_TRU_F64 : VOPC_64 <0x0000006f, "V_CMPS_TRU_F64", []>;
+defm V_CMPSX_F_F64 : VOPC_64 <0x00000070, "V_CMPSX_F_F64", []>;
+defm V_CMPSX_LT_F64 : VOPC_64 <0x00000071, "V_CMPSX_LT_F64", []>;
+defm V_CMPSX_EQ_F64 : VOPC_64 <0x00000072, "V_CMPSX_EQ_F64", []>;
+defm V_CMPSX_LE_F64 : VOPC_64 <0x00000073, "V_CMPSX_LE_F64", []>;
+defm V_CMPSX_GT_F64 : VOPC_64 <0x00000074, "V_CMPSX_GT_F64", []>;
+defm V_CMPSX_LG_F64 : VOPC_64 <0x00000075, "V_CMPSX_LG_F64", []>;
+defm V_CMPSX_GE_F64 : VOPC_64 <0x00000076, "V_CMPSX_GE_F64", []>;
+defm V_CMPSX_O_F64 : VOPC_64 <0x00000077, "V_CMPSX_O_F64", []>;
+defm V_CMPSX_U_F64 : VOPC_64 <0x00000078, "V_CMPSX_U_F64", []>;
+defm V_CMPSX_NGE_F64 : VOPC_64 <0x00000079, "V_CMPSX_NGE_F64", []>;
+defm V_CMPSX_NLG_F64 : VOPC_64 <0x0000007a, "V_CMPSX_NLG_F64", []>;
+defm V_CMPSX_NGT_F64 : VOPC_64 <0x0000007b, "V_CMPSX_NGT_F64", []>;
+defm V_CMPSX_NLE_F64 : VOPC_64 <0x0000007c, "V_CMPSX_NLE_F64", []>;
+defm V_CMPSX_NEQ_F64 : VOPC_64 <0x0000007d, "V_CMPSX_NEQ_F64", []>;
+defm V_CMPSX_NLT_F64 : VOPC_64 <0x0000007e, "V_CMPSX_NLT_F64", []>;
+defm V_CMPSX_TRU_F64 : VOPC_64 <0x0000007f, "V_CMPSX_TRU_F64", []>;
+defm V_CMP_F_I32 : VOPC_32 <0x00000080, "V_CMP_F_I32", []>;
+defm V_CMP_LT_I32 : VOPC_32 <0x00000081, "V_CMP_LT_I32", []>;
+def : Pat <
+  (i1 (setcc (i32 AllReg_32:$src0), VReg_32:$src1, COND_LT)),
+  (V_CMP_LT_I32_e64 AllReg_32:$src0, VReg_32:$src1)
+>;
+defm V_CMP_EQ_I32 : VOPC_32 <0x00000082, "V_CMP_EQ_I32", []>;
+def : Pat <
+  (i1 (setcc (i32 AllReg_32:$src0), VReg_32:$src1, COND_EQ)),
+  (V_CMP_EQ_I32_e64 AllReg_32:$src0, VReg_32:$src1)
+>;
+defm V_CMP_LE_I32 : VOPC_32 <0x00000083, "V_CMP_LE_I32", []>;
+def : Pat <
+  (i1 (setcc (i32 AllReg_32:$src0), VReg_32:$src1, COND_LE)),
+  (V_CMP_LE_I32_e64 AllReg_32:$src0, VReg_32:$src1)
+>;
+defm V_CMP_GT_I32 : VOPC_32 <0x00000084, "V_CMP_GT_I32", []>;
+def : Pat <
+  (i1 (setcc (i32 AllReg_32:$src0), VReg_32:$src1, COND_GT)),
+  (V_CMP_GT_I32_e64 AllReg_32:$src0, VReg_32:$src1)
+>;
+defm V_CMP_NE_I32 : VOPC_32 <0x00000085, "V_CMP_NE_I32", []>;
+def : Pat <
+  (i1 (setcc (i32 AllReg_32:$src0), VReg_32:$src1, COND_NE)),
+  (V_CMP_NE_I32_e64 AllReg_32:$src0, VReg_32:$src1)
+>;
+defm V_CMP_GE_I32 : VOPC_32 <0x00000086, "V_CMP_GE_I32", []>;
+def : Pat <
+  (i1 (setcc (i32 AllReg_32:$src0), VReg_32:$src1, COND_GE)),
+  (V_CMP_GE_I32_e64 AllReg_32:$src0, VReg_32:$src1)
+>;
+defm V_CMP_T_I32 : VOPC_32 <0x00000087, "V_CMP_T_I32", []>;
+
+let hasSideEffects = 1 in {
+
+defm V_CMPX_F_I32 : VOPC_32 <0x00000090, "V_CMPX_F_I32", []>;
+defm V_CMPX_LT_I32 : VOPC_32 <0x00000091, "V_CMPX_LT_I32", []>;
+defm V_CMPX_EQ_I32 : VOPC_32 <0x00000092, "V_CMPX_EQ_I32", []>;
+defm V_CMPX_LE_I32 : VOPC_32 <0x00000093, "V_CMPX_LE_I32", []>;
+defm V_CMPX_GT_I32 : VOPC_32 <0x00000094, "V_CMPX_GT_I32", []>;
+defm V_CMPX_NE_I32 : VOPC_32 <0x00000095, "V_CMPX_NE_I32", []>;
+defm V_CMPX_GE_I32 : VOPC_32 <0x00000096, "V_CMPX_GE_I32", []>;
+defm V_CMPX_T_I32 : VOPC_32 <0x00000097, "V_CMPX_T_I32", []>;
+
+} // End hasSideEffects
+
+defm V_CMP_F_I64 : VOPC_64 <0x000000a0, "V_CMP_F_I64", []>;
+defm V_CMP_LT_I64 : VOPC_64 <0x000000a1, "V_CMP_LT_I64", []>;
+defm V_CMP_EQ_I64 : VOPC_64 <0x000000a2, "V_CMP_EQ_I64", []>;
+defm V_CMP_LE_I64 : VOPC_64 <0x000000a3, "V_CMP_LE_I64", []>;
+defm V_CMP_GT_I64 : VOPC_64 <0x000000a4, "V_CMP_GT_I64", []>;
+defm V_CMP_NE_I64 : VOPC_64 <0x000000a5, "V_CMP_NE_I64", []>;
+defm V_CMP_GE_I64 : VOPC_64 <0x000000a6, "V_CMP_GE_I64", []>;
+defm V_CMP_T_I64 : VOPC_64 <0x000000a7, "V_CMP_T_I64", []>;
+
+let hasSideEffects = 1 in {
+
+defm V_CMPX_F_I64 : VOPC_64 <0x000000b0, "V_CMPX_F_I64", []>;
+defm V_CMPX_LT_I64 : VOPC_64 <0x000000b1, "V_CMPX_LT_I64", []>;
+defm V_CMPX_EQ_I64 : VOPC_64 <0x000000b2, "V_CMPX_EQ_I64", []>;
+defm V_CMPX_LE_I64 : VOPC_64 <0x000000b3, "V_CMPX_LE_I64", []>;
+defm V_CMPX_GT_I64 : VOPC_64 <0x000000b4, "V_CMPX_GT_I64", []>;
+defm V_CMPX_NE_I64 : VOPC_64 <0x000000b5, "V_CMPX_NE_I64", []>;
+defm V_CMPX_GE_I64 : VOPC_64 <0x000000b6, "V_CMPX_GE_I64", []>;
+defm V_CMPX_T_I64 : VOPC_64 <0x000000b7, "V_CMPX_T_I64", []>;
+
+} // End hasSideEffects
+
+defm V_CMP_F_U32 : VOPC_32 <0x000000c0, "V_CMP_F_U32", []>;
+defm V_CMP_LT_U32 : VOPC_32 <0x000000c1, "V_CMP_LT_U32", []>;
+defm V_CMP_EQ_U32 : VOPC_32 <0x000000c2, "V_CMP_EQ_U32", []>;
+defm V_CMP_LE_U32 : VOPC_32 <0x000000c3, "V_CMP_LE_U32", []>;
+defm V_CMP_GT_U32 : VOPC_32 <0x000000c4, "V_CMP_GT_U32", []>;
+defm V_CMP_NE_U32 : VOPC_32 <0x000000c5, "V_CMP_NE_U32", []>;
+defm V_CMP_GE_U32 : VOPC_32 <0x000000c6, "V_CMP_GE_U32", []>;
+defm V_CMP_T_U32 : VOPC_32 <0x000000c7, "V_CMP_T_U32", []>;
+
+let hasSideEffects = 1 in {
+
+defm V_CMPX_F_U32 : VOPC_32 <0x000000d0, "V_CMPX_F_U32", []>;
+defm V_CMPX_LT_U32 : VOPC_32 <0x000000d1, "V_CMPX_LT_U32", []>;
+defm V_CMPX_EQ_U32 : VOPC_32 <0x000000d2, "V_CMPX_EQ_U32", []>;
+defm V_CMPX_LE_U32 : VOPC_32 <0x000000d3, "V_CMPX_LE_U32", []>;
+defm V_CMPX_GT_U32 : VOPC_32 <0x000000d4, "V_CMPX_GT_U32", []>;
+defm V_CMPX_NE_U32 : VOPC_32 <0x000000d5, "V_CMPX_NE_U32", []>;
+defm V_CMPX_GE_U32 : VOPC_32 <0x000000d6, "V_CMPX_GE_U32", []>;
+defm V_CMPX_T_U32 : VOPC_32 <0x000000d7, "V_CMPX_T_U32", []>;
+
+} // End hasSideEffects
+
+defm V_CMP_F_U64 : VOPC_64 <0x000000e0, "V_CMP_F_U64", []>;
+defm V_CMP_LT_U64 : VOPC_64 <0x000000e1, "V_CMP_LT_U64", []>;
+defm V_CMP_EQ_U64 : VOPC_64 <0x000000e2, "V_CMP_EQ_U64", []>;
+defm V_CMP_LE_U64 : VOPC_64 <0x000000e3, "V_CMP_LE_U64", []>;
+defm V_CMP_GT_U64 : VOPC_64 <0x000000e4, "V_CMP_GT_U64", []>;
+defm V_CMP_NE_U64 : VOPC_64 <0x000000e5, "V_CMP_NE_U64", []>;
+defm V_CMP_GE_U64 : VOPC_64 <0x000000e6, "V_CMP_GE_U64", []>;
+defm V_CMP_T_U64 : VOPC_64 <0x000000e7, "V_CMP_T_U64", []>;
+defm V_CMPX_F_U64 : VOPC_64 <0x000000f0, "V_CMPX_F_U64", []>;
+defm V_CMPX_LT_U64 : VOPC_64 <0x000000f1, "V_CMPX_LT_U64", []>;
+defm V_CMPX_EQ_U64 : VOPC_64 <0x000000f2, "V_CMPX_EQ_U64", []>;
+defm V_CMPX_LE_U64 : VOPC_64 <0x000000f3, "V_CMPX_LE_U64", []>;
+defm V_CMPX_GT_U64 : VOPC_64 <0x000000f4, "V_CMPX_GT_U64", []>;
+defm V_CMPX_NE_U64 : VOPC_64 <0x000000f5, "V_CMPX_NE_U64", []>;
+defm V_CMPX_GE_U64 : VOPC_64 <0x000000f6, "V_CMPX_GE_U64", []>;
+defm V_CMPX_T_U64 : VOPC_64 <0x000000f7, "V_CMPX_T_U64", []>;
+defm V_CMP_CLASS_F32 : VOPC_32 <0x00000088, "V_CMP_CLASS_F32", []>;
+defm V_CMPX_CLASS_F32 : VOPC_32 <0x00000098, "V_CMPX_CLASS_F32", []>;
+defm V_CMP_CLASS_F64 : VOPC_64 <0x000000a8, "V_CMP_CLASS_F64", []>;
+defm V_CMPX_CLASS_F64 : VOPC_64 <0x000000b8, "V_CMPX_CLASS_F64", []>;
+//def BUFFER_LOAD_FORMAT_X : MUBUF_ <0x00000000, "BUFFER_LOAD_FORMAT_X", []>;
+//def BUFFER_LOAD_FORMAT_XY : MUBUF_ <0x00000001, "BUFFER_LOAD_FORMAT_XY", []>;
+//def BUFFER_LOAD_FORMAT_XYZ : MUBUF_ <0x00000002, "BUFFER_LOAD_FORMAT_XYZ", []>;
+def BUFFER_LOAD_FORMAT_XYZW : MUBUF_Load_Helper <0x00000003, "BUFFER_LOAD_FORMAT_XYZW", VReg_128>;
+//def BUFFER_STORE_FORMAT_X : MUBUF_ <0x00000004, "BUFFER_STORE_FORMAT_X", []>;
+//def BUFFER_STORE_FORMAT_XY : MUBUF_ <0x00000005, "BUFFER_STORE_FORMAT_XY", []>;
+//def BUFFER_STORE_FORMAT_XYZ : MUBUF_ <0x00000006, "BUFFER_STORE_FORMAT_XYZ", []>;
+//def BUFFER_STORE_FORMAT_XYZW : MUBUF_ <0x00000007, "BUFFER_STORE_FORMAT_XYZW", []>;
+//def BUFFER_LOAD_UBYTE : MUBUF_ <0x00000008, "BUFFER_LOAD_UBYTE", []>;
+//def BUFFER_LOAD_SBYTE : MUBUF_ <0x00000009, "BUFFER_LOAD_SBYTE", []>;
+//def BUFFER_LOAD_USHORT : MUBUF_ <0x0000000a, "BUFFER_LOAD_USHORT", []>;
+//def BUFFER_LOAD_SSHORT : MUBUF_ <0x0000000b, "BUFFER_LOAD_SSHORT", []>;
+//def BUFFER_LOAD_DWORD : MUBUF_ <0x0000000c, "BUFFER_LOAD_DWORD", []>;
+//def BUFFER_LOAD_DWORDX2 : MUBUF_DWORDX2 <0x0000000d, "BUFFER_LOAD_DWORDX2", []>;
+//def BUFFER_LOAD_DWORDX4 : MUBUF_DWORDX4 <0x0000000e, "BUFFER_LOAD_DWORDX4", []>;
+//def BUFFER_STORE_BYTE : MUBUF_ <0x00000018, "BUFFER_STORE_BYTE", []>;
+//def BUFFER_STORE_SHORT : MUBUF_ <0x0000001a, "BUFFER_STORE_SHORT", []>;
+//def BUFFER_STORE_DWORD : MUBUF_ <0x0000001c, "BUFFER_STORE_DWORD", []>;
+//def BUFFER_STORE_DWORDX2 : MUBUF_DWORDX2 <0x0000001d, "BUFFER_STORE_DWORDX2", []>;
+//def BUFFER_STORE_DWORDX4 : MUBUF_DWORDX4 <0x0000001e, "BUFFER_STORE_DWORDX4", []>;
+//def BUFFER_ATOMIC_SWAP : MUBUF_ <0x00000030, "BUFFER_ATOMIC_SWAP", []>;
+//def BUFFER_ATOMIC_CMPSWAP : MUBUF_ <0x00000031, "BUFFER_ATOMIC_CMPSWAP", []>;
+//def BUFFER_ATOMIC_ADD : MUBUF_ <0x00000032, "BUFFER_ATOMIC_ADD", []>;
+//def BUFFER_ATOMIC_SUB : MUBUF_ <0x00000033, "BUFFER_ATOMIC_SUB", []>;
+//def BUFFER_ATOMIC_RSUB : MUBUF_ <0x00000034, "BUFFER_ATOMIC_RSUB", []>;
+//def BUFFER_ATOMIC_SMIN : MUBUF_ <0x00000035, "BUFFER_ATOMIC_SMIN", []>;
+//def BUFFER_ATOMIC_UMIN : MUBUF_ <0x00000036, "BUFFER_ATOMIC_UMIN", []>;
+//def BUFFER_ATOMIC_SMAX : MUBUF_ <0x00000037, "BUFFER_ATOMIC_SMAX", []>;
+//def BUFFER_ATOMIC_UMAX : MUBUF_ <0x00000038, "BUFFER_ATOMIC_UMAX", []>;
+//def BUFFER_ATOMIC_AND : MUBUF_ <0x00000039, "BUFFER_ATOMIC_AND", []>;
+//def BUFFER_ATOMIC_OR : MUBUF_ <0x0000003a, "BUFFER_ATOMIC_OR", []>;
+//def BUFFER_ATOMIC_XOR : MUBUF_ <0x0000003b, "BUFFER_ATOMIC_XOR", []>;
+//def BUFFER_ATOMIC_INC : MUBUF_ <0x0000003c, "BUFFER_ATOMIC_INC", []>;
+//def BUFFER_ATOMIC_DEC : MUBUF_ <0x0000003d, "BUFFER_ATOMIC_DEC", []>;
+//def BUFFER_ATOMIC_FCMPSWAP : MUBUF_ <0x0000003e, "BUFFER_ATOMIC_FCMPSWAP", []>;
+//def BUFFER_ATOMIC_FMIN : MUBUF_ <0x0000003f, "BUFFER_ATOMIC_FMIN", []>;
+//def BUFFER_ATOMIC_FMAX : MUBUF_ <0x00000040, "BUFFER_ATOMIC_FMAX", []>;
+//def BUFFER_ATOMIC_SWAP_X2 : MUBUF_X2 <0x00000050, "BUFFER_ATOMIC_SWAP_X2", []>;
+//def BUFFER_ATOMIC_CMPSWAP_X2 : MUBUF_X2 <0x00000051, "BUFFER_ATOMIC_CMPSWAP_X2", []>;
+//def BUFFER_ATOMIC_ADD_X2 : MUBUF_X2 <0x00000052, "BUFFER_ATOMIC_ADD_X2", []>;
+//def BUFFER_ATOMIC_SUB_X2 : MUBUF_X2 <0x00000053, "BUFFER_ATOMIC_SUB_X2", []>;
+//def BUFFER_ATOMIC_RSUB_X2 : MUBUF_X2 <0x00000054, "BUFFER_ATOMIC_RSUB_X2", []>;
+//def BUFFER_ATOMIC_SMIN_X2 : MUBUF_X2 <0x00000055, "BUFFER_ATOMIC_SMIN_X2", []>;
+//def BUFFER_ATOMIC_UMIN_X2 : MUBUF_X2 <0x00000056, "BUFFER_ATOMIC_UMIN_X2", []>;
+//def BUFFER_ATOMIC_SMAX_X2 : MUBUF_X2 <0x00000057, "BUFFER_ATOMIC_SMAX_X2", []>;
+//def BUFFER_ATOMIC_UMAX_X2 : MUBUF_X2 <0x00000058, "BUFFER_ATOMIC_UMAX_X2", []>;
+//def BUFFER_ATOMIC_AND_X2 : MUBUF_X2 <0x00000059, "BUFFER_ATOMIC_AND_X2", []>;
+//def BUFFER_ATOMIC_OR_X2 : MUBUF_X2 <0x0000005a, "BUFFER_ATOMIC_OR_X2", []>;
+//def BUFFER_ATOMIC_XOR_X2 : MUBUF_X2 <0x0000005b, "BUFFER_ATOMIC_XOR_X2", []>;
+//def BUFFER_ATOMIC_INC_X2 : MUBUF_X2 <0x0000005c, "BUFFER_ATOMIC_INC_X2", []>;
+//def BUFFER_ATOMIC_DEC_X2 : MUBUF_X2 <0x0000005d, "BUFFER_ATOMIC_DEC_X2", []>;
+//def BUFFER_ATOMIC_FCMPSWAP_X2 : MUBUF_X2 <0x0000005e, "BUFFER_ATOMIC_FCMPSWAP_X2", []>;
+//def BUFFER_ATOMIC_FMIN_X2 : MUBUF_X2 <0x0000005f, "BUFFER_ATOMIC_FMIN_X2", []>;
+//def BUFFER_ATOMIC_FMAX_X2 : MUBUF_X2 <0x00000060, "BUFFER_ATOMIC_FMAX_X2", []>;
+//def BUFFER_WBINVL1_SC : MUBUF_WBINVL1 <0x00000070, "BUFFER_WBINVL1_SC", []>;
+//def BUFFER_WBINVL1 : MUBUF_WBINVL1 <0x00000071, "BUFFER_WBINVL1", []>;
+//def TBUFFER_LOAD_FORMAT_X : MTBUF_ <0x00000000, "TBUFFER_LOAD_FORMAT_X", []>;
+//def TBUFFER_LOAD_FORMAT_XY : MTBUF_ <0x00000001, "TBUFFER_LOAD_FORMAT_XY", []>;
+//def TBUFFER_LOAD_FORMAT_XYZ : MTBUF_ <0x00000002, "TBUFFER_LOAD_FORMAT_XYZ", []>;
+def TBUFFER_LOAD_FORMAT_XYZW : MTBUF_Load_Helper <0x00000003, "TBUFFER_LOAD_FORMAT_XYZW", VReg_128>;
+//def TBUFFER_STORE_FORMAT_X : MTBUF_ <0x00000004, "TBUFFER_STORE_FORMAT_X", []>;
+//def TBUFFER_STORE_FORMAT_XY : MTBUF_ <0x00000005, "TBUFFER_STORE_FORMAT_XY", []>;
+//def TBUFFER_STORE_FORMAT_XYZ : MTBUF_ <0x00000006, "TBUFFER_STORE_FORMAT_XYZ", []>;
+//def TBUFFER_STORE_FORMAT_XYZW : MTBUF_ <0x00000007, "TBUFFER_STORE_FORMAT_XYZW", []>;
+
+defm S_LOAD_DWORD : SMRD_32 <0x00000000, "S_LOAD_DWORD", SReg_32>;
+
+//def S_LOAD_DWORDX2 : SMRD_DWORDX2 <0x00000001, "S_LOAD_DWORDX2", []>;
+defm S_LOAD_DWORDX4 : SMRD_Helper <0x00000002, "S_LOAD_DWORDX4", SReg_128, v4i32>;
+defm S_LOAD_DWORDX8 : SMRD_Helper <0x00000003, "S_LOAD_DWORDX8", SReg_256, v8i32>;
+//def S_LOAD_DWORDX16 : SMRD_DWORDX16 <0x00000004, "S_LOAD_DWORDX16", []>;
+//def S_BUFFER_LOAD_DWORD : SMRD_ <0x00000008, "S_BUFFER_LOAD_DWORD", []>;
+//def S_BUFFER_LOAD_DWORDX2 : SMRD_DWORDX2 <0x00000009, "S_BUFFER_LOAD_DWORDX2", []>;
+//def S_BUFFER_LOAD_DWORDX4 : SMRD_DWORDX4 <0x0000000a, "S_BUFFER_LOAD_DWORDX4", []>;
+//def S_BUFFER_LOAD_DWORDX8 : SMRD_DWORDX8 <0x0000000b, "S_BUFFER_LOAD_DWORDX8", []>;
+//def S_BUFFER_LOAD_DWORDX16 : SMRD_DWORDX16 <0x0000000c, "S_BUFFER_LOAD_DWORDX16", []>;
+
+//def S_MEMTIME : SMRD_ <0x0000001e, "S_MEMTIME", []>;
+//def S_DCACHE_INV : SMRD_ <0x0000001f, "S_DCACHE_INV", []>;
+//def IMAGE_LOAD : MIMG_NoPattern_ <"IMAGE_LOAD", 0x00000000>;
+//def IMAGE_LOAD_MIP : MIMG_NoPattern_ <"IMAGE_LOAD_MIP", 0x00000001>;
+//def IMAGE_LOAD_PCK : MIMG_NoPattern_ <"IMAGE_LOAD_PCK", 0x00000002>;
+//def IMAGE_LOAD_PCK_SGN : MIMG_NoPattern_ <"IMAGE_LOAD_PCK_SGN", 0x00000003>;
+//def IMAGE_LOAD_MIP_PCK : MIMG_NoPattern_ <"IMAGE_LOAD_MIP_PCK", 0x00000004>;
+//def IMAGE_LOAD_MIP_PCK_SGN : MIMG_NoPattern_ <"IMAGE_LOAD_MIP_PCK_SGN", 0x00000005>;
+//def IMAGE_STORE : MIMG_NoPattern_ <"IMAGE_STORE", 0x00000008>;
+//def IMAGE_STORE_MIP : MIMG_NoPattern_ <"IMAGE_STORE_MIP", 0x00000009>;
+//def IMAGE_STORE_PCK : MIMG_NoPattern_ <"IMAGE_STORE_PCK", 0x0000000a>;
+//def IMAGE_STORE_MIP_PCK : MIMG_NoPattern_ <"IMAGE_STORE_MIP_PCK", 0x0000000b>;
+//def IMAGE_GET_RESINFO : MIMG_NoPattern_ <"IMAGE_GET_RESINFO", 0x0000000e>;
+//def IMAGE_ATOMIC_SWAP : MIMG_NoPattern_ <"IMAGE_ATOMIC_SWAP", 0x0000000f>;
+//def IMAGE_ATOMIC_CMPSWAP : MIMG_NoPattern_ <"IMAGE_ATOMIC_CMPSWAP", 0x00000010>;
+//def IMAGE_ATOMIC_ADD : MIMG_NoPattern_ <"IMAGE_ATOMIC_ADD", 0x00000011>;
+//def IMAGE_ATOMIC_SUB : MIMG_NoPattern_ <"IMAGE_ATOMIC_SUB", 0x00000012>;
+//def IMAGE_ATOMIC_RSUB : MIMG_NoPattern_ <"IMAGE_ATOMIC_RSUB", 0x00000013>;
+//def IMAGE_ATOMIC_SMIN : MIMG_NoPattern_ <"IMAGE_ATOMIC_SMIN", 0x00000014>;
+//def IMAGE_ATOMIC_UMIN : MIMG_NoPattern_ <"IMAGE_ATOMIC_UMIN", 0x00000015>;
+//def IMAGE_ATOMIC_SMAX : MIMG_NoPattern_ <"IMAGE_ATOMIC_SMAX", 0x00000016>;
+//def IMAGE_ATOMIC_UMAX : MIMG_NoPattern_ <"IMAGE_ATOMIC_UMAX", 0x00000017>;
+//def IMAGE_ATOMIC_AND : MIMG_NoPattern_ <"IMAGE_ATOMIC_AND", 0x00000018>;
+//def IMAGE_ATOMIC_OR : MIMG_NoPattern_ <"IMAGE_ATOMIC_OR", 0x00000019>;
+//def IMAGE_ATOMIC_XOR : MIMG_NoPattern_ <"IMAGE_ATOMIC_XOR", 0x0000001a>;
+//def IMAGE_ATOMIC_INC : MIMG_NoPattern_ <"IMAGE_ATOMIC_INC", 0x0000001b>;
+//def IMAGE_ATOMIC_DEC : MIMG_NoPattern_ <"IMAGE_ATOMIC_DEC", 0x0000001c>;
+//def IMAGE_ATOMIC_FCMPSWAP : MIMG_NoPattern_ <"IMAGE_ATOMIC_FCMPSWAP", 0x0000001d>;
+//def IMAGE_ATOMIC_FMIN : MIMG_NoPattern_ <"IMAGE_ATOMIC_FMIN", 0x0000001e>;
+//def IMAGE_ATOMIC_FMAX : MIMG_NoPattern_ <"IMAGE_ATOMIC_FMAX", 0x0000001f>;
+def IMAGE_SAMPLE : MIMG_Load_Helper <0x00000020, "IMAGE_SAMPLE">; 
+//def IMAGE_SAMPLE_CL : MIMG_NoPattern_ <"IMAGE_SAMPLE_CL", 0x00000021>;
+def IMAGE_SAMPLE_D : MIMG_Load_Helper <0x00000022, "IMAGE_SAMPLE_D">;
+//def IMAGE_SAMPLE_D_CL : MIMG_NoPattern_ <"IMAGE_SAMPLE_D_CL", 0x00000023>;
+def IMAGE_SAMPLE_L : MIMG_Load_Helper <0x00000024, "IMAGE_SAMPLE_L">;
+def IMAGE_SAMPLE_B : MIMG_Load_Helper <0x00000025, "IMAGE_SAMPLE_B">;
+//def IMAGE_SAMPLE_B_CL : MIMG_NoPattern_ <"IMAGE_SAMPLE_B_CL", 0x00000026>;
+//def IMAGE_SAMPLE_LZ : MIMG_NoPattern_ <"IMAGE_SAMPLE_LZ", 0x00000027>;
+//def IMAGE_SAMPLE_C : MIMG_NoPattern_ <"IMAGE_SAMPLE_C", 0x00000028>;
+//def IMAGE_SAMPLE_C_CL : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_CL", 0x00000029>;
+//def IMAGE_SAMPLE_C_D : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_D", 0x0000002a>;
+//def IMAGE_SAMPLE_C_D_CL : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_D_CL", 0x0000002b>;
+//def IMAGE_SAMPLE_C_L : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_L", 0x0000002c>;
+//def IMAGE_SAMPLE_C_B : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_B", 0x0000002d>;
+//def IMAGE_SAMPLE_C_B_CL : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_B_CL", 0x0000002e>;
+//def IMAGE_SAMPLE_C_LZ : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_LZ", 0x0000002f>;
+//def IMAGE_SAMPLE_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_O", 0x00000030>;
+//def IMAGE_SAMPLE_CL_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_CL_O", 0x00000031>;
+//def IMAGE_SAMPLE_D_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_D_O", 0x00000032>;
+//def IMAGE_SAMPLE_D_CL_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_D_CL_O", 0x00000033>;
+//def IMAGE_SAMPLE_L_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_L_O", 0x00000034>;
+//def IMAGE_SAMPLE_B_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_B_O", 0x00000035>;
+//def IMAGE_SAMPLE_B_CL_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_B_CL_O", 0x00000036>;
+//def IMAGE_SAMPLE_LZ_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_LZ_O", 0x00000037>;
+//def IMAGE_SAMPLE_C_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_O", 0x00000038>;
+//def IMAGE_SAMPLE_C_CL_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_CL_O", 0x00000039>;
+//def IMAGE_SAMPLE_C_D_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_D_O", 0x0000003a>;
+//def IMAGE_SAMPLE_C_D_CL_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_D_CL_O", 0x0000003b>;
+//def IMAGE_SAMPLE_C_L_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_L_O", 0x0000003c>;
+//def IMAGE_SAMPLE_C_B_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_B_O", 0x0000003d>;
+//def IMAGE_SAMPLE_C_B_CL_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_B_CL_O", 0x0000003e>;
+//def IMAGE_SAMPLE_C_LZ_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_LZ_O", 0x0000003f>;
+//def IMAGE_GATHER4 : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4", 0x00000040>;
+//def IMAGE_GATHER4_CL : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_CL", 0x00000041>;
+//def IMAGE_GATHER4_L : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_L", 0x00000044>;
+//def IMAGE_GATHER4_B : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_B", 0x00000045>;
+//def IMAGE_GATHER4_B_CL : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_B_CL", 0x00000046>;
+//def IMAGE_GATHER4_LZ : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_LZ", 0x00000047>;
+//def IMAGE_GATHER4_C : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_C", 0x00000048>;
+//def IMAGE_GATHER4_C_CL : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_C_CL", 0x00000049>;
+//def IMAGE_GATHER4_C_L : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_C_L", 0x0000004c>;
+//def IMAGE_GATHER4_C_B : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_C_B", 0x0000004d>;
+//def IMAGE_GATHER4_C_B_CL : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_C_B_CL", 0x0000004e>;
+//def IMAGE_GATHER4_C_LZ : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_C_LZ", 0x0000004f>;
+//def IMAGE_GATHER4_O : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_O", 0x00000050>;
+//def IMAGE_GATHER4_CL_O : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_CL_O", 0x00000051>;
+//def IMAGE_GATHER4_L_O : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_L_O", 0x00000054>;
+//def IMAGE_GATHER4_B_O : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_B_O", 0x00000055>;
+//def IMAGE_GATHER4_B_CL_O : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_B_CL_O", 0x00000056>;
+//def IMAGE_GATHER4_LZ_O : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_LZ_O", 0x00000057>;
+//def IMAGE_GATHER4_C_O : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_C_O", 0x00000058>;
+//def IMAGE_GATHER4_C_CL_O : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_C_CL_O", 0x00000059>;
+//def IMAGE_GATHER4_C_L_O : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_C_L_O", 0x0000005c>;
+//def IMAGE_GATHER4_C_B_O : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_C_B_O", 0x0000005d>;
+//def IMAGE_GATHER4_C_B_CL_O : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_C_B_CL_O", 0x0000005e>;
+//def IMAGE_GATHER4_C_LZ_O : MIMG_NoPattern_GATHER4 <"IMAGE_GATHER4_C_LZ_O", 0x0000005f>;
+//def IMAGE_GET_LOD : MIMG_NoPattern_ <"IMAGE_GET_LOD", 0x00000060>;
+//def IMAGE_SAMPLE_CD : MIMG_NoPattern_ <"IMAGE_SAMPLE_CD", 0x00000068>;
+//def IMAGE_SAMPLE_CD_CL : MIMG_NoPattern_ <"IMAGE_SAMPLE_CD_CL", 0x00000069>;
+//def IMAGE_SAMPLE_C_CD : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_CD", 0x0000006a>;
+//def IMAGE_SAMPLE_C_CD_CL : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_CD_CL", 0x0000006b>;
+//def IMAGE_SAMPLE_CD_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_CD_O", 0x0000006c>;
+//def IMAGE_SAMPLE_CD_CL_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_CD_CL_O", 0x0000006d>;
+//def IMAGE_SAMPLE_C_CD_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_CD_O", 0x0000006e>;
+//def IMAGE_SAMPLE_C_CD_CL_O : MIMG_NoPattern_ <"IMAGE_SAMPLE_C_CD_CL_O", 0x0000006f>;
+//def IMAGE_RSRC256 : MIMG_NoPattern_RSRC256 <"IMAGE_RSRC256", 0x0000007e>;
+//def IMAGE_SAMPLER : MIMG_NoPattern_ <"IMAGE_SAMPLER", 0x0000007f>;
+//def V_NOP : VOP1_ <0x00000000, "V_NOP", []>;
+
+let neverHasSideEffects = 1 in {
+defm V_MOV_B32 : VOP1_32 <0x00000001, "V_MOV_B32", []>;
+}  // End neverHasSideEffects
+defm V_READFIRSTLANE_B32 : VOP1_32 <0x00000002, "V_READFIRSTLANE_B32", []>;
+//defm V_CVT_I32_F64 : VOP1_32 <0x00000003, "V_CVT_I32_F64", []>;
+//defm V_CVT_F64_I32 : VOP1_64 <0x00000004, "V_CVT_F64_I32", []>;
+defm V_CVT_F32_I32 : VOP1_32 <0x00000005, "V_CVT_F32_I32",
+  [(set VReg_32:$dst, (sint_to_fp AllReg_32:$src0))]
+>;
+//defm V_CVT_F32_U32 : VOP1_32 <0x00000006, "V_CVT_F32_U32", []>;
+//defm V_CVT_U32_F32 : VOP1_32 <0x00000007, "V_CVT_U32_F32", []>;
+defm V_CVT_I32_F32 : VOP1_32 <0x00000008, "V_CVT_I32_F32",
+  [(set VReg_32:$dst, (fp_to_sint AllReg_32:$src0))]
+>;
+defm V_MOV_FED_B32 : VOP1_32 <0x00000009, "V_MOV_FED_B32", []>;
+////def V_CVT_F16_F32 : VOP1_F16 <0x0000000a, "V_CVT_F16_F32", []>;
+//defm V_CVT_F32_F16 : VOP1_32 <0x0000000b, "V_CVT_F32_F16", []>;
+//defm V_CVT_RPI_I32_F32 : VOP1_32 <0x0000000c, "V_CVT_RPI_I32_F32", []>;
+//defm V_CVT_FLR_I32_F32 : VOP1_32 <0x0000000d, "V_CVT_FLR_I32_F32", []>;
+//defm V_CVT_OFF_F32_I4 : VOP1_32 <0x0000000e, "V_CVT_OFF_F32_I4", []>;
+//defm V_CVT_F32_F64 : VOP1_32 <0x0000000f, "V_CVT_F32_F64", []>;
+//defm V_CVT_F64_F32 : VOP1_64 <0x00000010, "V_CVT_F64_F32", []>;
+//defm V_CVT_F32_UBYTE0 : VOP1_32 <0x00000011, "V_CVT_F32_UBYTE0", []>;
+//defm V_CVT_F32_UBYTE1 : VOP1_32 <0x00000012, "V_CVT_F32_UBYTE1", []>;
+//defm V_CVT_F32_UBYTE2 : VOP1_32 <0x00000013, "V_CVT_F32_UBYTE2", []>;
+//defm V_CVT_F32_UBYTE3 : VOP1_32 <0x00000014, "V_CVT_F32_UBYTE3", []>;
+//defm V_CVT_U32_F64 : VOP1_32 <0x00000015, "V_CVT_U32_F64", []>;
+//defm V_CVT_F64_U32 : VOP1_64 <0x00000016, "V_CVT_F64_U32", []>;
+defm V_FRACT_F32 : VOP1_32 <0x00000020, "V_FRACT_F32",
+  [(set VReg_32:$dst, (AMDGPUfract AllReg_32:$src0))]
+>;
+defm V_TRUNC_F32 : VOP1_32 <0x00000021, "V_TRUNC_F32", []>;
+defm V_CEIL_F32 : VOP1_32 <0x00000022, "V_CEIL_F32", []>;
+defm V_RNDNE_F32 : VOP1_32 <0x00000023, "V_RNDNE_F32",
+  [(set VReg_32:$dst, (frint AllReg_32:$src0))]
+>;
+defm V_FLOOR_F32 : VOP1_32 <0x00000024, "V_FLOOR_F32",
+  [(set VReg_32:$dst, (ffloor AllReg_32:$src0))]
+>;
+defm V_EXP_F32 : VOP1_32 <0x00000025, "V_EXP_F32",
+  [(set VReg_32:$dst, (fexp2 AllReg_32:$src0))]
+>;
+defm V_LOG_CLAMP_F32 : VOP1_32 <0x00000026, "V_LOG_CLAMP_F32", []>;
+defm V_LOG_F32 : VOP1_32 <0x00000027, "V_LOG_F32", []>;
+defm V_RCP_CLAMP_F32 : VOP1_32 <0x00000028, "V_RCP_CLAMP_F32", []>;
+defm V_RCP_LEGACY_F32 : VOP1_32 <0x00000029, "V_RCP_LEGACY_F32", []>;
+defm V_RCP_F32 : VOP1_32 <0x0000002a, "V_RCP_F32",
+  [(set VReg_32:$dst, (fdiv FP_ONE, AllReg_32:$src0))]
+>;
+defm V_RCP_IFLAG_F32 : VOP1_32 <0x0000002b, "V_RCP_IFLAG_F32", []>;
+defm V_RSQ_CLAMP_F32 : VOP1_32 <0x0000002c, "V_RSQ_CLAMP_F32", []>;
+defm V_RSQ_LEGACY_F32 : VOP1_32 <
+  0x0000002d, "V_RSQ_LEGACY_F32",
+  [(set VReg_32:$dst, (int_AMDGPU_rsq AllReg_32:$src0))]
+>;
+defm V_RSQ_F32 : VOP1_32 <0x0000002e, "V_RSQ_F32", []>;
+defm V_RCP_F64 : VOP1_64 <0x0000002f, "V_RCP_F64", []>;
+defm V_RCP_CLAMP_F64 : VOP1_64 <0x00000030, "V_RCP_CLAMP_F64", []>;
+defm V_RSQ_F64 : VOP1_64 <0x00000031, "V_RSQ_F64", []>;
+defm V_RSQ_CLAMP_F64 : VOP1_64 <0x00000032, "V_RSQ_CLAMP_F64", []>;
+defm V_SQRT_F32 : VOP1_32 <0x00000033, "V_SQRT_F32", []>;
+defm V_SQRT_F64 : VOP1_64 <0x00000034, "V_SQRT_F64", []>;
+defm V_SIN_F32 : VOP1_32 <0x00000035, "V_SIN_F32", []>;
+defm V_COS_F32 : VOP1_32 <0x00000036, "V_COS_F32", []>;
+defm V_NOT_B32 : VOP1_32 <0x00000037, "V_NOT_B32", []>;
+defm V_BFREV_B32 : VOP1_32 <0x00000038, "V_BFREV_B32", []>;
+defm V_FFBH_U32 : VOP1_32 <0x00000039, "V_FFBH_U32", []>;
+defm V_FFBL_B32 : VOP1_32 <0x0000003a, "V_FFBL_B32", []>;
+defm V_FFBH_I32 : VOP1_32 <0x0000003b, "V_FFBH_I32", []>;
+//defm V_FREXP_EXP_I32_F64 : VOP1_32 <0x0000003c, "V_FREXP_EXP_I32_F64", []>;
+defm V_FREXP_MANT_F64 : VOP1_64 <0x0000003d, "V_FREXP_MANT_F64", []>;
+defm V_FRACT_F64 : VOP1_64 <0x0000003e, "V_FRACT_F64", []>;
+//defm V_FREXP_EXP_I32_F32 : VOP1_32 <0x0000003f, "V_FREXP_EXP_I32_F32", []>;
+defm V_FREXP_MANT_F32 : VOP1_32 <0x00000040, "V_FREXP_MANT_F32", []>;
+//def V_CLREXCP : VOP1_ <0x00000041, "V_CLREXCP", []>;
+defm V_MOVRELD_B32 : VOP1_32 <0x00000042, "V_MOVRELD_B32", []>;
+defm V_MOVRELS_B32 : VOP1_32 <0x00000043, "V_MOVRELS_B32", []>;
+defm V_MOVRELSD_B32 : VOP1_32 <0x00000044, "V_MOVRELSD_B32", []>;
+
+def V_INTERP_P1_F32 : VINTRP <
+  0x00000000,
+  (outs VReg_32:$dst),
+  (ins VReg_32:$i, i32imm:$attr_chan, i32imm:$attr, M0Reg:$m0),
+  "V_INTERP_P1_F32",
+  []> {
+  let DisableEncoding = "$m0";
+}
+
+def V_INTERP_P2_F32 : VINTRP <
+  0x00000001,
+  (outs VReg_32:$dst),
+  (ins VReg_32:$src0, VReg_32:$j, i32imm:$attr_chan, i32imm:$attr, M0Reg:$m0),
+  "V_INTERP_P2_F32",
+  []> {
+
+  let Constraints = "$src0 = $dst";
+  let DisableEncoding = "$src0,$m0";
+
+}
+
+def V_INTERP_MOV_F32 : VINTRP <
+  0x00000002,
+  (outs VReg_32:$dst),
+  (ins i32imm:$attr_chan, i32imm:$attr, M0Reg:$m0),
+  "V_INTERP_MOV_F32",
+  []> {
+  let VSRC = 0;
+  let DisableEncoding = "$m0";
+}
+
+//def S_NOP : SOPP_ <0x00000000, "S_NOP", []>;
+
+let isTerminator = 1 in {
+
+def S_ENDPGM : SOPP <0x00000001, (ins), "S_ENDPGM",
+  [(IL_retflag)]> {
+  let SIMM16 = 0;
+  let isBarrier = 1;
+  let hasCtrlDep = 1;
+}
+
+let isBranch = 1 in {
+def S_BRANCH : SOPP <
+  0x00000002, (ins brtarget:$target), "S_BRANCH",
+  []
+>;
+
+let DisableEncoding = "$scc" in {
+def S_CBRANCH_SCC0 : SOPP <
+  0x00000004, (ins brtarget:$target, SCCReg:$scc),
+  "S_CBRANCH_SCC0", []
+>;
+def S_CBRANCH_SCC1 : SOPP <
+  0x00000005, (ins brtarget:$target, SCCReg:$scc),
+  "S_CBRANCH_SCC1",
+  []
+>;
+} // End DisableEncoding = "$scc"
+
+def S_CBRANCH_VCCZ : SOPP <
+  0x00000006, (ins brtarget:$target, VCCReg:$vcc),
+  "S_CBRANCH_VCCZ",
+  []
+>;
+def S_CBRANCH_VCCNZ : SOPP <
+  0x00000007, (ins brtarget:$target, VCCReg:$vcc),
+  "S_CBRANCH_VCCNZ",
+  []
+>;
+
+let DisableEncoding = "$exec" in {
+def S_CBRANCH_EXECZ : SOPP <
+  0x00000008, (ins brtarget:$target, EXECReg:$exec),
+  "S_CBRANCH_EXECZ",
+  []
+>;
+def S_CBRANCH_EXECNZ : SOPP <
+  0x00000009, (ins brtarget:$target, EXECReg:$exec),
+  "S_CBRANCH_EXECNZ",
+  []
+>;
+} // End DisableEncoding = "$exec"
+
+
+} // End isBranch = 1
+} // End isTerminator = 1
+
+//def S_BARRIER : SOPP_ <0x0000000a, "S_BARRIER", []>;
+let hasSideEffects = 1 in {
+def S_WAITCNT : SOPP <0x0000000c, (ins i32imm:$simm16), "S_WAITCNT $simm16",
+  []
+>;
+} // End hasSideEffects
+//def S_SETHALT : SOPP_ <0x0000000d, "S_SETHALT", []>;
+//def S_SLEEP : SOPP_ <0x0000000e, "S_SLEEP", []>;
+//def S_SETPRIO : SOPP_ <0x0000000f, "S_SETPRIO", []>;
+//def S_SENDMSG : SOPP_ <0x00000010, "S_SENDMSG", []>;
+//def S_SENDMSGHALT : SOPP_ <0x00000011, "S_SENDMSGHALT", []>;
+//def S_TRAP : SOPP_ <0x00000012, "S_TRAP", []>;
+//def S_ICACHE_INV : SOPP_ <0x00000013, "S_ICACHE_INV", []>;
+//def S_INCPERFLEVEL : SOPP_ <0x00000014, "S_INCPERFLEVEL", []>;
+//def S_DECPERFLEVEL : SOPP_ <0x00000015, "S_DECPERFLEVEL", []>;
+//def S_TTRACEDATA : SOPP_ <0x00000016, "S_TTRACEDATA", []>;
+
+def V_CNDMASK_B32_e32 : VOP2 <0x00000000, (outs VReg_32:$dst),
+  (ins AllReg_32:$src0, VReg_32:$src1, VCCReg:$vcc), "V_CNDMASK_B32_e32",
+  []
+>{
+  let DisableEncoding = "$vcc";
+}
+
+def V_CNDMASK_B32_e64 : VOP3 <0x00000100, (outs VReg_32:$dst),
+  (ins VReg_32:$src0, VReg_32:$src1, SReg_1:$src2, InstFlag:$abs, InstFlag:$clamp, InstFlag:$omod, InstFlag:$neg),
+  "V_CNDMASK_B32_e64",
+  [(set (i32 VReg_32:$dst), (select SReg_1:$src2, VReg_32:$src1, VReg_32:$src0))]
+>;
+
+//f32 pattern for V_CNDMASK_B32_e64
+def : Pat <
+  (f32 (select SReg_1:$src2, VReg_32:$src1, VReg_32:$src0)),
+  (V_CNDMASK_B32_e64 VReg_32:$src0, VReg_32:$src1, SReg_1:$src2)
+>;
+
+defm V_READLANE_B32 : VOP2_32 <0x00000001, "V_READLANE_B32", []>;
+defm V_WRITELANE_B32 : VOP2_32 <0x00000002, "V_WRITELANE_B32", []>;
+
+defm V_ADD_F32 : VOP2_32 <0x00000003, "V_ADD_F32", []>;
+def : Pat <
+  (f32 (fadd AllReg_32:$src0, VReg_32:$src1)),
+  (V_ADD_F32_e32  AllReg_32:$src0, VReg_32:$src1)
+>;
+
+defm V_SUB_F32 : VOP2_32 <0x00000004, "V_SUB_F32", []>;
+def : Pat <
+  (f32 (fsub AllReg_32:$src0, VReg_32:$src1)),
+  (V_SUB_F32_e32  AllReg_32:$src0, VReg_32:$src1)
+>;
+defm V_SUBREV_F32 : VOP2_32 <0x00000005, "V_SUBREV_F32", []>;
+defm V_MAC_LEGACY_F32 : VOP2_32 <0x00000006, "V_MAC_LEGACY_F32", []>;
+defm V_MUL_LEGACY_F32 : VOP2_32 <
+  0x00000007, "V_MUL_LEGACY_F32",
+  [(set VReg_32:$dst, (int_AMDGPU_mul AllReg_32:$src0, VReg_32:$src1))]
+>;
+
+defm V_MUL_F32 : VOP2_32 <0x00000008, "V_MUL_F32",
+  [(set VReg_32:$dst, (fmul AllReg_32:$src0, VReg_32:$src1))]
+>;
+//defm V_MUL_I32_I24 : VOP2_32 <0x00000009, "V_MUL_I32_I24", []>;
+//defm V_MUL_HI_I32_I24 : VOP2_32 <0x0000000a, "V_MUL_HI_I32_I24", []>;
+//defm V_MUL_U32_U24 : VOP2_32 <0x0000000b, "V_MUL_U32_U24", []>;
+//defm V_MUL_HI_U32_U24 : VOP2_32 <0x0000000c, "V_MUL_HI_U32_U24", []>;
+defm V_MIN_LEGACY_F32 : VOP2_32 <0x0000000d, "V_MIN_LEGACY_F32",
+  [(set VReg_32:$dst, (AMDGPUfmin AllReg_32:$src0, VReg_32:$src1))]
+>;
+
+defm V_MAX_LEGACY_F32 : VOP2_32 <0x0000000e, "V_MAX_LEGACY_F32",
+  [(set VReg_32:$dst, (AMDGPUfmax AllReg_32:$src0, VReg_32:$src1))]
+>;
+defm V_MIN_F32 : VOP2_32 <0x0000000f, "V_MIN_F32", []>;
+defm V_MAX_F32 : VOP2_32 <0x00000010, "V_MAX_F32", []>;
+defm V_MIN_I32 : VOP2_32 <0x00000011, "V_MIN_I32", []>;
+defm V_MAX_I32 : VOP2_32 <0x00000012, "V_MAX_I32", []>;
+defm V_MIN_U32 : VOP2_32 <0x00000013, "V_MIN_U32", []>;
+defm V_MAX_U32 : VOP2_32 <0x00000014, "V_MAX_U32", []>;
+defm V_LSHR_B32 : VOP2_32 <0x00000015, "V_LSHR_B32", []>;
+defm V_LSHRREV_B32 : VOP2_32 <0x00000016, "V_LSHRREV_B32", []>;
+defm V_ASHR_I32 : VOP2_32 <0x00000017, "V_ASHR_I32", []>;
+defm V_ASHRREV_I32 : VOP2_32 <0x00000018, "V_ASHRREV_I32", []>;
+defm V_LSHL_B32 : VOP2_32 <0x00000019, "V_LSHL_B32", []>;
+defm V_LSHLREV_B32 : VOP2_32 <0x0000001a, "V_LSHLREV_B32", []>;
+defm V_AND_B32 : VOP2_32 <0x0000001b, "V_AND_B32",
+  [(set VReg_32:$dst, (and AllReg_32:$src0, VReg_32:$src1))]
+>;
+defm V_OR_B32 : VOP2_32 <0x0000001c, "V_OR_B32",
+  [(set VReg_32:$dst, (or AllReg_32:$src0, VReg_32:$src1))]
+>;
+defm V_XOR_B32 : VOP2_32 <0x0000001d, "V_XOR_B32",
+  [(set VReg_32:$dst, (xor AllReg_32:$src0, VReg_32:$src1))]
+>;
+defm V_BFM_B32 : VOP2_32 <0x0000001e, "V_BFM_B32", []>;
+defm V_MAC_F32 : VOP2_32 <0x0000001f, "V_MAC_F32", []>;
+defm V_MADMK_F32 : VOP2_32 <0x00000020, "V_MADMK_F32", []>;
+defm V_MADAK_F32 : VOP2_32 <0x00000021, "V_MADAK_F32", []>;
+//defm V_BCNT_U32_B32 : VOP2_32 <0x00000022, "V_BCNT_U32_B32", []>;
+//defm V_MBCNT_LO_U32_B32 : VOP2_32 <0x00000023, "V_MBCNT_LO_U32_B32", []>;
+//defm V_MBCNT_HI_U32_B32 : VOP2_32 <0x00000024, "V_MBCNT_HI_U32_B32", []>;
+let Defs = [VCC] in { // Carry-out goes to VCC
+defm V_ADD_I32 : VOP2_32 <0x00000025, "V_ADD_I32",
+  [(set VReg_32:$dst, (add (i32 AllReg_32:$src0), (i32 VReg_32:$src1)))]
+>;
+defm V_SUB_I32 : VOP2_32 <0x00000026, "V_SUB_I32",
+  [(set VReg_32:$dst, (sub (i32 AllReg_32:$src0), (i32 VReg_32:$src1)))]
+>;
+} // End Defs = [VCC]
+defm V_SUBREV_I32 : VOP2_32 <0x00000027, "V_SUBREV_I32", []>;
+defm V_ADDC_U32 : VOP2_32 <0x00000028, "V_ADDC_U32", []>;
+defm V_SUBB_U32 : VOP2_32 <0x00000029, "V_SUBB_U32", []>;
+defm V_SUBBREV_U32 : VOP2_32 <0x0000002a, "V_SUBBREV_U32", []>;
+defm V_LDEXP_F32 : VOP2_32 <0x0000002b, "V_LDEXP_F32", []>;
+////def V_CVT_PKACCUM_U8_F32 : VOP2_U8 <0x0000002c, "V_CVT_PKACCUM_U8_F32", []>;
+////def V_CVT_PKNORM_I16_F32 : VOP2_I16 <0x0000002d, "V_CVT_PKNORM_I16_F32", []>;
+////def V_CVT_PKNORM_U16_F32 : VOP2_U16 <0x0000002e, "V_CVT_PKNORM_U16_F32", []>;
+defm V_CVT_PKRTZ_F16_F32 : VOP2_32 <0x0000002f, "V_CVT_PKRTZ_F16_F32",
+ [(set VReg_32:$dst, (int_SI_packf16 AllReg_32:$src0, VReg_32:$src1))]
+>;
+////def V_CVT_PK_U16_U32 : VOP2_U16 <0x00000030, "V_CVT_PK_U16_U32", []>;
+////def V_CVT_PK_I16_I32 : VOP2_I16 <0x00000031, "V_CVT_PK_I16_I32", []>;
+def S_CMP_EQ_I32 : SOPC_32 <0x00000000, "S_CMP_EQ_I32", []>;
+def S_CMP_LG_I32 : SOPC_32 <0x00000001, "S_CMP_LG_I32", []>;
+def S_CMP_GT_I32 : SOPC_32 <0x00000002, "S_CMP_GT_I32", []>;
+def S_CMP_GE_I32 : SOPC_32 <0x00000003, "S_CMP_GE_I32", []>;
+def S_CMP_LT_I32 : SOPC_32 <0x00000004, "S_CMP_LT_I32", []>;
+def S_CMP_LE_I32 : SOPC_32 <0x00000005, "S_CMP_LE_I32", []>;
+def S_CMP_EQ_U32 : SOPC_32 <0x00000006, "S_CMP_EQ_U32", []>;
+def S_CMP_LG_U32 : SOPC_32 <0x00000007, "S_CMP_LG_U32", []>;
+def S_CMP_GT_U32 : SOPC_32 <0x00000008, "S_CMP_GT_U32", []>;
+def S_CMP_GE_U32 : SOPC_32 <0x00000009, "S_CMP_GE_U32", []>;
+def S_CMP_LT_U32 : SOPC_32 <0x0000000a, "S_CMP_LT_U32", []>;
+def S_CMP_LE_U32 : SOPC_32 <0x0000000b, "S_CMP_LE_U32", []>;
+////def S_BITCMP0_B32 : SOPC_BITCMP0 <0x0000000c, "S_BITCMP0_B32", []>;
+////def S_BITCMP1_B32 : SOPC_BITCMP1 <0x0000000d, "S_BITCMP1_B32", []>;
+////def S_BITCMP0_B64 : SOPC_BITCMP0 <0x0000000e, "S_BITCMP0_B64", []>;
+////def S_BITCMP1_B64 : SOPC_BITCMP1 <0x0000000f, "S_BITCMP1_B64", []>;
+//def S_SETVSKIP : SOPC_ <0x00000010, "S_SETVSKIP", []>;
+
+let neverHasSideEffects = 1 in {
+
+def V_MAD_LEGACY_F32 : VOP3_32 <0x00000140, "V_MAD_LEGACY_F32", []>;
+def V_MAD_F32 : VOP3_32 <0x00000141, "V_MAD_F32", []>;
+//def V_MAD_I32_I24 : VOP3_32 <0x00000142, "V_MAD_I32_I24", []>;
+//def V_MAD_U32_U24 : VOP3_32 <0x00000143, "V_MAD_U32_U24", []>;
+
+} // End neverHasSideEffects
+def V_CUBEID_F32 : VOP3_32 <0x00000144, "V_CUBEID_F32", []>;
+def V_CUBESC_F32 : VOP3_32 <0x00000145, "V_CUBESC_F32", []>;
+def V_CUBETC_F32 : VOP3_32 <0x00000146, "V_CUBETC_F32", []>;
+def V_CUBEMA_F32 : VOP3_32 <0x00000147, "V_CUBEMA_F32", []>;
+def V_BFE_U32 : VOP3_32 <0x00000148, "V_BFE_U32", []>;
+def V_BFE_I32 : VOP3_32 <0x00000149, "V_BFE_I32", []>;
+def V_BFI_B32 : VOP3_32 <0x0000014a, "V_BFI_B32", []>;
+def V_FMA_F32 : VOP3_32 <0x0000014b, "V_FMA_F32", []>;
+def V_FMA_F64 : VOP3_64 <0x0000014c, "V_FMA_F64", []>;
+//def V_LERP_U8 : VOP3_U8 <0x0000014d, "V_LERP_U8", []>;
+def V_ALIGNBIT_B32 : VOP3_32 <0x0000014e, "V_ALIGNBIT_B32", []>;
+def V_ALIGNBYTE_B32 : VOP3_32 <0x0000014f, "V_ALIGNBYTE_B32", []>;
+def V_MULLIT_F32 : VOP3_32 <0x00000150, "V_MULLIT_F32", []>;
+////def V_MIN3_F32 : VOP3_MIN3 <0x00000151, "V_MIN3_F32", []>;
+////def V_MIN3_I32 : VOP3_MIN3 <0x00000152, "V_MIN3_I32", []>;
+////def V_MIN3_U32 : VOP3_MIN3 <0x00000153, "V_MIN3_U32", []>;
+////def V_MAX3_F32 : VOP3_MAX3 <0x00000154, "V_MAX3_F32", []>;
+////def V_MAX3_I32 : VOP3_MAX3 <0x00000155, "V_MAX3_I32", []>;
+////def V_MAX3_U32 : VOP3_MAX3 <0x00000156, "V_MAX3_U32", []>;
+////def V_MED3_F32 : VOP3_MED3 <0x00000157, "V_MED3_F32", []>;
+////def V_MED3_I32 : VOP3_MED3 <0x00000158, "V_MED3_I32", []>;
+////def V_MED3_U32 : VOP3_MED3 <0x00000159, "V_MED3_U32", []>;
+//def V_SAD_U8 : VOP3_U8 <0x0000015a, "V_SAD_U8", []>;
+//def V_SAD_HI_U8 : VOP3_U8 <0x0000015b, "V_SAD_HI_U8", []>;
+//def V_SAD_U16 : VOP3_U16 <0x0000015c, "V_SAD_U16", []>;
+def V_SAD_U32 : VOP3_32 <0x0000015d, "V_SAD_U32", []>;
+////def V_CVT_PK_U8_F32 : VOP3_U8 <0x0000015e, "V_CVT_PK_U8_F32", []>;
+def V_DIV_FIXUP_F32 : VOP3_32 <0x0000015f, "V_DIV_FIXUP_F32", []>;
+def V_DIV_FIXUP_F64 : VOP3_64 <0x00000160, "V_DIV_FIXUP_F64", []>;
+def V_LSHL_B64 : VOP3_64 <0x00000161, "V_LSHL_B64", []>;
+def V_LSHR_B64 : VOP3_64 <0x00000162, "V_LSHR_B64", []>;
+def V_ASHR_I64 : VOP3_64 <0x00000163, "V_ASHR_I64", []>;
+def V_ADD_F64 : VOP3_64 <0x00000164, "V_ADD_F64", []>;
+def V_MUL_F64 : VOP3_64 <0x00000165, "V_MUL_F64", []>;
+def V_MIN_F64 : VOP3_64 <0x00000166, "V_MIN_F64", []>;
+def V_MAX_F64 : VOP3_64 <0x00000167, "V_MAX_F64", []>;
+def V_LDEXP_F64 : VOP3_64 <0x00000168, "V_LDEXP_F64", []>;
+def V_MUL_LO_U32 : VOP3_32 <0x00000169, "V_MUL_LO_U32", []>;
+def V_MUL_HI_U32 : VOP3_32 <0x0000016a, "V_MUL_HI_U32", []>;
+def V_MUL_LO_I32 : VOP3_32 <0x0000016b, "V_MUL_LO_I32", []>;
+def V_MUL_HI_I32 : VOP3_32 <0x0000016c, "V_MUL_HI_I32", []>;
+def V_DIV_SCALE_F32 : VOP3_32 <0x0000016d, "V_DIV_SCALE_F32", []>;
+def V_DIV_SCALE_F64 : VOP3_64 <0x0000016e, "V_DIV_SCALE_F64", []>;
+def V_DIV_FMAS_F32 : VOP3_32 <0x0000016f, "V_DIV_FMAS_F32", []>;
+def V_DIV_FMAS_F64 : VOP3_64 <0x00000170, "V_DIV_FMAS_F64", []>;
+//def V_MSAD_U8 : VOP3_U8 <0x00000171, "V_MSAD_U8", []>;
+//def V_QSAD_U8 : VOP3_U8 <0x00000172, "V_QSAD_U8", []>;
+//def V_MQSAD_U8 : VOP3_U8 <0x00000173, "V_MQSAD_U8", []>;
+def V_TRIG_PREOP_F64 : VOP3_64 <0x00000174, "V_TRIG_PREOP_F64", []>;
+def S_ADD_U32 : SOP2_32 <0x00000000, "S_ADD_U32", []>;
+def S_SUB_U32 : SOP2_32 <0x00000001, "S_SUB_U32", []>;
+def S_ADD_I32 : SOP2_32 <0x00000002, "S_ADD_I32", []>;
+def S_SUB_I32 : SOP2_32 <0x00000003, "S_SUB_I32", []>;
+def S_ADDC_U32 : SOP2_32 <0x00000004, "S_ADDC_U32", []>;
+def S_SUBB_U32 : SOP2_32 <0x00000005, "S_SUBB_U32", []>;
+def S_MIN_I32 : SOP2_32 <0x00000006, "S_MIN_I32", []>;
+def S_MIN_U32 : SOP2_32 <0x00000007, "S_MIN_U32", []>;
+def S_MAX_I32 : SOP2_32 <0x00000008, "S_MAX_I32", []>;
+def S_MAX_U32 : SOP2_32 <0x00000009, "S_MAX_U32", []>;
+
+def S_CSELECT_B32 : SOP2 <
+  0x0000000a, (outs SReg_32:$dst),
+  (ins SReg_32:$src0, SReg_32:$src1, SCCReg:$scc), "S_CSELECT_B32",
+  [(set (i32 SReg_32:$dst), (select SCCReg:$scc, SReg_32:$src0, SReg_32:$src1))]
+>;
+
+def S_CSELECT_B64 : SOP2_64 <0x0000000b, "S_CSELECT_B64", []>;
+
+// f32 pattern for S_CSELECT_B32
+def : Pat <
+  (f32 (select SCCReg:$scc, SReg_32:$src0, SReg_32:$src1)),
+  (S_CSELECT_B32 SReg_32:$src0, SReg_32:$src1, SCCReg:$scc)
+>;
+
+def S_AND_B32 : SOP2_32 <0x0000000e, "S_AND_B32", []>;
+
+def S_AND_B64 : SOP2_64 <0x0000000f, "S_AND_B64",
+  [(set SReg_64:$dst, (and SReg_64:$src0, SReg_64:$src1))]
+>;
+def S_AND_VCC : SOP2_VCC <0x0000000f, "S_AND_B64",
+  [(set SReg_1:$vcc, (SIvcc_and SReg_64:$src0, SReg_64:$src1))]
+>;
+def S_OR_B32 : SOP2_32 <0x00000010, "S_OR_B32", []>;
+def S_OR_B64 : SOP2_64 <0x00000011, "S_OR_B64", []>;
+def S_XOR_B32 : SOP2_32 <0x00000012, "S_XOR_B32", []>;
+def S_XOR_B64 : SOP2_64 <0x00000013, "S_XOR_B64", []>;
+////def S_ANDN2_B32 : SOP2_ANDN2 <0x00000014, "S_ANDN2_B32", []>;
+////def S_ANDN2_B64 : SOP2_ANDN2 <0x00000015, "S_ANDN2_B64", []>;
+////def S_ORN2_B32 : SOP2_ORN2 <0x00000016, "S_ORN2_B32", []>;
+////def S_ORN2_B64 : SOP2_ORN2 <0x00000017, "S_ORN2_B64", []>;
+def S_NAND_B32 : SOP2_32 <0x00000018, "S_NAND_B32", []>;
+def S_NAND_B64 : SOP2_64 <0x00000019, "S_NAND_B64", []>;
+def S_NOR_B32 : SOP2_32 <0x0000001a, "S_NOR_B32", []>;
+def S_NOR_B64 : SOP2_64 <0x0000001b, "S_NOR_B64", []>;
+def S_XNOR_B32 : SOP2_32 <0x0000001c, "S_XNOR_B32", []>;
+def S_XNOR_B64 : SOP2_64 <0x0000001d, "S_XNOR_B64", []>;
+def S_LSHL_B32 : SOP2_32 <0x0000001e, "S_LSHL_B32", []>;
+def S_LSHL_B64 : SOP2_64 <0x0000001f, "S_LSHL_B64", []>;
+def S_LSHR_B32 : SOP2_32 <0x00000020, "S_LSHR_B32", []>;
+def S_LSHR_B64 : SOP2_64 <0x00000021, "S_LSHR_B64", []>;
+def S_ASHR_I32 : SOP2_32 <0x00000022, "S_ASHR_I32", []>;
+def S_ASHR_I64 : SOP2_64 <0x00000023, "S_ASHR_I64", []>;
+def S_BFM_B32 : SOP2_32 <0x00000024, "S_BFM_B32", []>;
+def S_BFM_B64 : SOP2_64 <0x00000025, "S_BFM_B64", []>;
+def S_MUL_I32 : SOP2_32 <0x00000026, "S_MUL_I32", []>;
+def S_BFE_U32 : SOP2_32 <0x00000027, "S_BFE_U32", []>;
+def S_BFE_I32 : SOP2_32 <0x00000028, "S_BFE_I32", []>;
+def S_BFE_U64 : SOP2_64 <0x00000029, "S_BFE_U64", []>;
+def S_BFE_I64 : SOP2_64 <0x0000002a, "S_BFE_I64", []>;
+//def S_CBRANCH_G_FORK : SOP2_ <0x0000002b, "S_CBRANCH_G_FORK", []>;
+def S_ABSDIFF_I32 : SOP2_32 <0x0000002c, "S_ABSDIFF_I32", []>;
+
+class V_MOV_IMM <Operand immType, SDNode immNode> : InstSI <
+  (outs VReg_32:$dst),
+  (ins immType:$src0),
+  "V_MOV_IMM",
+   [(set VReg_32:$dst, (immNode:$src0))]
+>;
+
+let isCodeGenOnly = 1, isPseudo = 1 in {
+
+def V_MOV_IMM_I32 : V_MOV_IMM<i32imm, imm>;
+def V_MOV_IMM_F32 : V_MOV_IMM<f32imm, fpimm>;
+
+def S_MOV_IMM_I32 : InstSI <
+  (outs SReg_32:$dst),
+  (ins i32imm:$src0),
+  "S_MOV_IMM_I32",
+  [(set SReg_32:$dst, (imm:$src0))]
+>;
+
+// i64 immediates aren't really supported in hardware, but LLVM will use the i64
+// type for indices on load and store instructions.  The pattern for
+// S_MOV_IMM_I64 will only match i64 immediates that can fit into 32-bits,
+// which the hardware can handle.
+def S_MOV_IMM_I64 : InstSI <
+  (outs SReg_64:$dst),
+  (ins i64imm:$src0),
+  "S_MOV_IMM_I64 $dst, $src0",
+  [(set SReg_64:$dst, (IMM32bitIn64bit:$src0))]
+>;
+
+} // End isCodeGenOnly, isPseudo = 1
+
+class SI_LOAD_LITERAL<Operand ImmType> :
+    Enc32 <(outs), (ins ImmType:$imm), "LOAD_LITERAL $imm", []> {
+
+  bits<32> imm;
+  let Inst{31-0} = imm;
+}
+
+def SI_LOAD_LITERAL_I32 : SI_LOAD_LITERAL<i32imm>;
+def SI_LOAD_LITERAL_F32 : SI_LOAD_LITERAL<f32imm>;
+
+let isCodeGenOnly = 1, isPseudo = 1 in {
+
+def SET_M0 : InstSI <
+  (outs SReg_32:$dst),
+  (ins i32imm:$src0),
+  "SET_M0",
+  [(set SReg_32:$dst, (int_SI_set_M0 imm:$src0))]
+>;
+
+def LOAD_CONST : AMDGPUShaderInst <
+  (outs GPRF32:$dst),
+  (ins i32imm:$src),
+  "LOAD_CONST $dst, $src",
+  [(set GPRF32:$dst, (int_AMDGPU_load_const imm:$src))]
+>;
+
+let usesCustomInserter = 1 in {
+
+def SI_V_CNDLT : InstSI <
+  (outs VReg_32:$dst),
+  (ins VReg_32:$src0, VReg_32:$src1, VReg_32:$src2),
+  "SI_V_CNDLT $dst, $src0, $src1, $src2",
+  [(set VReg_32:$dst, (int_AMDGPU_cndlt VReg_32:$src0, VReg_32:$src1, VReg_32:$src2))]
+>;
+
+def SI_INTERP : InstSI <
+  (outs VReg_32:$dst),
+  (ins VReg_32:$i, VReg_32:$j, i32imm:$attr_chan, i32imm:$attr, SReg_32:$params),
+  "SI_INTERP $dst, $i, $j, $attr_chan, $attr, $params",
+  []
+>;
+
+def SI_INTERP_CONST : InstSI <
+  (outs VReg_32:$dst),
+  (ins i32imm:$attr_chan, i32imm:$attr, SReg_32:$params),
+  "SI_INTERP_CONST $dst, $attr_chan, $attr, $params",
+  [(set VReg_32:$dst, (int_SI_fs_interp_constant imm:$attr_chan,
+                                                 imm:$attr, SReg_32:$params))]
+>;
+
+def SI_KIL : InstSI <
+  (outs),
+  (ins VReg_32:$src),
+  "SI_KIL $src",
+  [(int_AMDGPU_kill VReg_32:$src)]
+>;
+
+def SI_WQM : InstSI <
+  (outs),
+  (ins),
+  "SI_WQM",
+  [(int_SI_wqm)]
+>;
+
+} // end usesCustomInserter 
+
+// SI Psuedo branch instructions.  These are used by the CFG structurizer pass
+// and should be lowered to ISA instructions prior to codegen.
+
+let isBranch = 1, isTerminator = 1, mayLoad = 0, mayStore = 0,
+                                                 hasSideEffects = 0 in {
+def SI_IF_NZ : InstSI <
+  (outs),
+  (ins brtarget:$target, SReg_1:$vcc),
+  "SI_BRANCH_NZ",
+  [(IL_brcond bb:$target, SReg_1:$vcc)]
+>;
+
+def SI_IF_Z : InstSI <
+  (outs),
+  (ins brtarget:$target, SReg_1:$vcc),
+  "SI_BRANCH_Z",
+  []
+>;
+} // end isBranch = 1, isTerminator = 1, mayLoad = 0, mayStore = 0,
+  //     hasSideEffects = 0
+} // end IsCodeGenOnly, isPseudo
+
+/* int_SI_vs_load_input */
+def : Pat<
+  (int_SI_vs_load_input SReg_128:$tlst, IMM12bit:$attr_offset,
+                        VReg_32:$buf_idx_vgpr),
+  (BUFFER_LOAD_FORMAT_XYZW imm:$attr_offset, 0, 1, 0, 0, 0,
+                           VReg_32:$buf_idx_vgpr, SReg_128:$tlst,
+                           0, 0, (i32 SREG_LIT_0))
+>;
+
+/* int_SI_export */
+def : Pat <
+  (int_SI_export imm:$en, imm:$vm, imm:$done, imm:$tgt, imm:$compr,
+                 VReg_32:$src0,VReg_32:$src1, VReg_32:$src2, VReg_32:$src3),
+  (EXP imm:$en, imm:$tgt, imm:$compr, imm:$done, imm:$vm,
+       VReg_32:$src0, VReg_32:$src1, VReg_32:$src2, VReg_32:$src3)
+>;
+
+/* int_SI_sample */
+def : Pat <
+  (int_SI_sample imm:$writemask, VReg_128:$coord, SReg_256:$rsrc, SReg_128:$sampler),
+  (IMAGE_SAMPLE imm:$writemask, 0, 0, 0, 0, 0, 0, 0, VReg_128:$coord,
+                SReg_256:$rsrc, SReg_128:$sampler)
+>;
+
+/* int_SI_sample_lod */
+def : Pat <
+  (int_SI_sample_lod imm:$writemask, VReg_128:$coord, SReg_256:$rsrc, SReg_128:$sampler),
+  (IMAGE_SAMPLE_L imm:$writemask, 0, 0, 0, 0, 0, 0, 0, VReg_128:$coord,
+                  SReg_256:$rsrc, SReg_128:$sampler)
+>;
+
+/* int_SI_sample_bias */
+def : Pat <
+  (int_SI_sample_bias imm:$writemask, VReg_128:$coord, SReg_256:$rsrc, SReg_128:$sampler),
+  (IMAGE_SAMPLE_B imm:$writemask, 0, 0, 0, 0, 0, 0, 0, VReg_128:$coord,
+                  SReg_256:$rsrc, SReg_128:$sampler)
+>;
+
+def CLAMP_SI : CLAMP<VReg_32>;
+def FABS_SI : FABS<VReg_32>;
+def FNEG_SI : FNEG<VReg_32>;
+
+def : Extract_Element <f32, v4f32, VReg_128, 0, sel_x>;
+def : Extract_Element <f32, v4f32, VReg_128, 1, sel_y>;
+def : Extract_Element <f32, v4f32, VReg_128, 2, sel_z>;
+def : Extract_Element <f32, v4f32, VReg_128, 3, sel_w>;
+
+def : Insert_Element <f32, v4f32, VReg_32, VReg_128, 4, sel_x>;
+def : Insert_Element <f32, v4f32, VReg_32, VReg_128, 5, sel_y>;
+def : Insert_Element <f32, v4f32, VReg_32, VReg_128, 6, sel_z>;
+def : Insert_Element <f32, v4f32, VReg_32, VReg_128, 7, sel_w>;
+
+def : Vector_Build <v4f32, VReg_128, f32, VReg_32>;
+def : Vector_Build <v4i32, SReg_128, i32, SReg_32>;
+
+def : BitConvert <i32, f32, SReg_32>;
+def : BitConvert <i32, f32, VReg_32>;
+
+def : BitConvert <f32, i32, SReg_32>;
+def : BitConvert <f32, i32, VReg_32>;
+
+def : Pat <
+  (i64 (SIsreg1_bitcast SReg_1:$vcc)),
+  (S_MOV_B64 (COPY_TO_REGCLASS SReg_1:$vcc, SReg_64))
+>;
+
+def : Pat <
+  (i1 (SIsreg1_bitcast SReg_64:$vcc)),
+  (COPY_TO_REGCLASS SReg_64:$vcc, SReg_1)
+>;
+
+def : Pat <
+  (i64 (SIvcc_bitcast VCCReg:$vcc)),
+  (S_MOV_B64 (COPY_TO_REGCLASS VCCReg:$vcc, SReg_64))
+>;
+
+def : Pat <
+  (i1 (SIvcc_bitcast SReg_64:$vcc)),
+  (COPY_TO_REGCLASS SReg_64:$vcc, VCCReg)
+>;
+
+/********** ===================== **********/
+/********** Interpolation Paterns **********/
+/********** ===================== **********/
+
+def : Pat <
+  (int_SI_fs_interp_linear_center imm:$attr_chan, imm:$attr, SReg_32:$params),
+  (SI_INTERP (f32 LINEAR_CENTER_I), (f32 LINEAR_CENTER_J), imm:$attr_chan,
+             imm:$attr, SReg_32:$params)
+>;
+
+def : Pat <
+  (int_SI_fs_interp_linear_centroid imm:$attr_chan, imm:$attr, SReg_32:$params),
+  (SI_INTERP (f32 LINEAR_CENTROID_I), (f32 LINEAR_CENTROID_J), imm:$attr_chan,
+             imm:$attr, SReg_32:$params)
+>;
+
+def : Pat <
+  (int_SI_fs_interp_persp_center imm:$attr_chan, imm:$attr, SReg_32:$params),
+  (SI_INTERP (f32 PERSP_CENTER_I), (f32 PERSP_CENTER_J), imm:$attr_chan,
+             imm:$attr, SReg_32:$params)
+>;
+
+def : Pat <
+  (int_SI_fs_interp_persp_centroid imm:$attr_chan, imm:$attr, SReg_32:$params),
+  (SI_INTERP (f32 PERSP_CENTROID_I), (f32 PERSP_CENTROID_J), imm:$attr_chan,
+             imm:$attr, SReg_32:$params)
+>;
+
+def : Pat <
+  (int_SI_fs_read_face),
+  (f32 FRONT_FACE)
+>;
+
+def : Pat <
+  (int_SI_fs_read_pos 0),
+  (f32 POS_X_FLOAT)
+>;
+
+def : Pat <
+  (int_SI_fs_read_pos 1),
+  (f32 POS_Y_FLOAT)
+>;
+
+def : Pat <
+  (int_SI_fs_read_pos 2),
+  (f32 POS_Z_FLOAT)
+>;
+
+def : Pat <
+  (int_SI_fs_read_pos 3),
+  (f32 POS_W_FLOAT)
+>;
+
+/********** ================== **********/
+/********** Intrinsic Patterns **********/
+/********** ================== **********/
+
+/* llvm.AMDGPU.pow */
+/* XXX: We are using IEEE MUL, not the 0 * anything = 0 MUL, is this correct? */
+def : POW_Common <V_LOG_F32_e32, V_EXP_F32_e32, V_MUL_F32_e32, VReg_32>;
+
+def : Pat <
+  (int_AMDGPU_div AllReg_32:$src0, AllReg_32:$src1),
+  (V_MUL_LEGACY_F32_e32 AllReg_32:$src0, (V_RCP_LEGACY_F32_e32 AllReg_32:$src1))
+>;
+
+def : Pat<
+  (fdiv AllReg_32:$src0, AllReg_32:$src1),
+  (V_MUL_F32_e32 AllReg_32:$src0, (V_RCP_F32_e32 AllReg_32:$src1))
+>;
+
+def : Pat <
+  (int_AMDGPU_kilp),
+  (SI_KIL (V_MOV_IMM_I32 0xbf800000))
+>;
+
+def : Pat <
+  (int_AMDGPU_cube VReg_128:$src),
+  (INSERT_SUBREG (INSERT_SUBREG (INSERT_SUBREG (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)),
+    (V_CUBETC_F32 (EXTRACT_SUBREG VReg_128:$src, sel_x),
+                  (EXTRACT_SUBREG VReg_128:$src, sel_y),
+                  (EXTRACT_SUBREG VReg_128:$src, sel_z),
+                  0, 0, 0, 0), sel_x),
+    (V_CUBESC_F32 (EXTRACT_SUBREG VReg_128:$src, sel_x),
+                  (EXTRACT_SUBREG VReg_128:$src, sel_y),
+                  (EXTRACT_SUBREG VReg_128:$src, sel_z),
+                  0, 0, 0, 0), sel_y),
+    (V_CUBEMA_F32 (EXTRACT_SUBREG VReg_128:$src, sel_x),
+                  (EXTRACT_SUBREG VReg_128:$src, sel_y),
+                  (EXTRACT_SUBREG VReg_128:$src, sel_z),
+                  0, 0, 0, 0), sel_z),
+    (V_CUBEID_F32 (EXTRACT_SUBREG VReg_128:$src, sel_x),
+                  (EXTRACT_SUBREG VReg_128:$src, sel_y),
+                  (EXTRACT_SUBREG VReg_128:$src, sel_z),
+                  0, 0, 0, 0), sel_w)
+>;
+
+/********** ================== **********/
+/**********   VOP3 Patterns    **********/
+/********** ================== **********/
+
+def : Pat <(f32 (IL_mad AllReg_32:$src0, VReg_32:$src1, VReg_32:$src2)),
+           (V_MAD_LEGACY_F32 AllReg_32:$src0, VReg_32:$src1, VReg_32:$src2,
+            0, 0, 0, 0)>;
+
+} // End isSI predicate
diff --git a/lib/Target/R600/SIIntrinsics.td b/lib/Target/R600/SIIntrinsics.td
new file mode 100644
index 0000000000..1008fc42cc
--- /dev/null
+++ b/lib/Target/R600/SIIntrinsics.td
@@ -0,0 +1,42 @@
+//===-- SIIntrinsics.td - SI Intrinsic defs ----------------*- tablegen -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// SI Intrinsic Definitions
+//
+//===----------------------------------------------------------------------===//
+
+
+let TargetPrefix = "SI", isTarget = 1 in {
+
+  def int_SI_packf16 : Intrinsic <[llvm_i32_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
+  def int_SI_export : Intrinsic <[], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_i32_ty, llvm_float_ty, llvm_float_ty, llvm_float_ty, llvm_float_ty], []>;
+  /* XXX: We may need a seperate intrinsic here for loading integer values */
+  def int_SI_load_const : Intrinsic <[llvm_float_ty], [llvm_i64_ty, llvm_i32_ty], []>;
+  def int_SI_vs_load_buffer_index : Intrinsic <[llvm_i32_ty], [], [IntrNoMem]>;
+  def int_SI_vs_load_input : Intrinsic <[llvm_v4f32_ty], [llvm_v4i32_ty, llvm_i16_ty, llvm_i32_ty], [IntrReadMem]> ;
+  def int_SI_wqm : Intrinsic <[], [], []>;
+
+  def int_SI_sample : Intrinsic <[llvm_v4f32_ty], [llvm_i32_ty, llvm_v4f32_ty, llvm_v8i32_ty, llvm_v4i32_ty], [IntrReadMem]>;
+  def int_SI_sample_bias : Intrinsic <[llvm_v4f32_ty], [llvm_i32_ty, llvm_v4f32_ty, llvm_v8i32_ty, llvm_v4i32_ty], [IntrReadMem]>;
+  def int_SI_sample_lod : Intrinsic <[llvm_v4f32_ty], [llvm_i32_ty, llvm_v4f32_ty, llvm_v8i32_ty, llvm_v4i32_ty], [IntrReadMem]>;
+
+  /* Interpolation Intrinsics */
+
+  def int_SI_set_M0 : Intrinsic <[llvm_i32_ty], [llvm_i32_ty]>;
+  class Interp : Intrinsic <[llvm_float_ty], [llvm_i32_ty, llvm_i32_ty, llvm_i32_ty], [IntrReadMem]>;
+
+  def int_SI_fs_interp_linear_center : Interp;
+  def int_SI_fs_interp_linear_centroid : Interp;
+  def int_SI_fs_interp_persp_center : Interp;
+  def int_SI_fs_interp_persp_centroid : Interp;
+  def int_SI_fs_interp_constant : Interp;
+
+  def int_SI_fs_read_face : Intrinsic <[llvm_float_ty], [], [IntrNoMem]>;
+  def int_SI_fs_read_pos : Intrinsic <[llvm_float_ty], [llvm_i32_ty], [IntrNoMem]>;
+}
diff --git a/lib/Target/R600/SILowerControlFlow.cpp b/lib/Target/R600/SILowerControlFlow.cpp
new file mode 100644
index 0000000000..277b647f67
--- /dev/null
+++ b/lib/Target/R600/SILowerControlFlow.cpp
@@ -0,0 +1,191 @@
+//===-- SILowerControlFlow.cpp - Use predicates for control flow ----------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief This pass lowers the pseudo control flow instructions (SI_IF_NZ, ELSE, ENDIF)
+/// to predicated instructions.
+///
+/// All control flow (except loops) is handled using predicated instructions and
+/// a predicate stack.  Each Scalar ALU controls the operations of 64 Vector
+/// ALUs.  The Scalar ALU can update the predicate for any of the Vector ALUs
+/// by writting to the 64-bit EXEC register (each bit corresponds to a
+/// single vector ALU).  Typically, for predicates, a vector ALU will write
+/// to its bit of the VCC register (like EXEC VCC is 64-bits, one for each
+/// Vector ALU) and then the ScalarALU will AND the VCC register with the
+/// EXEC to update the predicates.
+///
+/// For example:
+/// %VCC = V_CMP_GT_F32 %VGPR1, %VGPR2
+/// SI_IF_NZ %VCC
+///   %VGPR0 = V_ADD_F32 %VGPR0, %VGPR0
+/// ELSE
+///   %VGPR0 = V_SUB_F32 %VGPR0, %VGPR0
+/// ENDIF
+///
+/// becomes:
+///
+/// %SGPR0 = S_AND_SAVEEXEC_B64 %VCC  // Save and update the exec mask
+/// %SGPR0 = S_XOR_B64 %SGPR0, %EXEC  // Clear live bits from saved exec mask
+/// S_CBRANCH_EXECZ label0            // This instruction is an
+///                                   // optimization which allows us to
+///                                   // branch if all the bits of
+///                                   // EXEC are zero.
+/// %VGPR0 = V_ADD_F32 %VGPR0, %VGPR0 // Do the IF block of the branch
+///
+/// label0:
+/// %SGPR0 = S_OR_SAVEEXEC_B64 %EXEC   // Restore the exec mask for the Then block
+/// %EXEC = S_XOR_B64 %SGPR0, %EXEC    // Clear live bits from saved exec mask
+/// S_BRANCH_EXECZ label1              // Use our branch optimization
+///                                    // instruction again.
+/// %VGPR0 = V_SUB_F32 %VGPR0, %VGPR   // Do the THEN block
+/// label1:
+/// %EXEC = S_OR_B64 %EXEC, %SGPR2     // Re-enable saved exec mask bits
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "SIInstrInfo.h"
+#include "SIMachineFunctionInfo.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+
+using namespace llvm;
+
+namespace {
+
+class SILowerControlFlowPass : public MachineFunctionPass {
+
+private:
+  static char ID;
+  const TargetInstrInfo *TII;
+  std::vector<unsigned> PredicateStack;
+  std::vector<unsigned> UnusedRegisters;
+
+  unsigned allocReg();
+  void freeReg(unsigned Reg);
+
+public:
+  SILowerControlFlowPass(TargetMachine &tm) :
+    MachineFunctionPass(ID), TII(tm.getInstrInfo()) { }
+
+  virtual bool runOnMachineFunction(MachineFunction &MF);
+
+  const char *getPassName() const {
+    return "SI Lower control flow instructions";
+  }
+
+};
+
+} // End anonymous namespace
+
+char SILowerControlFlowPass::ID = 0;
+
+FunctionPass *llvm::createSILowerControlFlowPass(TargetMachine &tm) {
+  return new SILowerControlFlowPass(tm);
+}
+
+bool SILowerControlFlowPass::runOnMachineFunction(MachineFunction &MF) {
+
+  // Find all the unused registers that can be used for the predicate stack.
+  for (TargetRegisterClass::iterator I = AMDGPU::SReg_64RegClass.begin(),
+                                     S = AMDGPU::SReg_64RegClass.end();
+                                     I != S; ++I) {
+    unsigned Reg = *I;
+    if (!MF.getRegInfo().isPhysRegUsed(Reg)) {
+      UnusedRegisters.insert(UnusedRegisters.begin(), Reg);
+    }
+  }
+
+  for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
+                                                  BB != BB_E; ++BB) {
+    MachineBasicBlock &MBB = *BB;
+    for (MachineBasicBlock::iterator I = MBB.begin(), Next = llvm::next(I);
+                               I != MBB.end(); I = Next) {
+      Next = llvm::next(I);
+      MachineInstr &MI = *I;
+      unsigned Reg;
+      switch (MI.getOpcode()) {
+        default: break;
+        case AMDGPU::SI_IF_NZ:
+          Reg = allocReg();
+          BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_AND_SAVEEXEC_B64),
+                  Reg)
+                  .addOperand(MI.getOperand(0)); // VCC
+          BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_XOR_B64),
+                  Reg)
+                  .addReg(Reg)
+                  .addReg(AMDGPU::EXEC);
+          MI.eraseFromParent();
+          PredicateStack.push_back(Reg);
+          break;
+
+        case AMDGPU::ELSE:
+          Reg = PredicateStack.back();
+          BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_OR_SAVEEXEC_B64),
+                  Reg)
+                  .addReg(Reg);
+          BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_XOR_B64),
+                  AMDGPU::EXEC)
+                  .addReg(Reg)
+                  .addReg(AMDGPU::EXEC);
+          MI.eraseFromParent();
+          break;
+
+        case AMDGPU::ENDIF:
+          Reg = PredicateStack.back();
+          PredicateStack.pop_back();
+          BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_OR_B64),
+                  AMDGPU::EXEC)
+                  .addReg(AMDGPU::EXEC)
+                  .addReg(Reg);
+          freeReg(Reg);
+
+          if (MF.getInfo<SIMachineFunctionInfo>()->ShaderType == ShaderType::PIXEL &&
+              PredicateStack.empty()) {
+            // If the exec mask is non-zero, skip the next two instructions
+            BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_CBRANCH_EXECNZ))
+                    .addImm(3)
+                    .addReg(AMDGPU::EXEC);
+
+            // Exec mask is zero: Export to NULL target...
+            BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::EXP))
+                    .addImm(0)
+                    .addImm(0x09) // V_008DFC_SQ_EXP_NULL
+                    .addImm(0)
+                    .addImm(1)
+                    .addImm(1)
+                    .addReg(AMDGPU::SREG_LIT_0)
+                    .addReg(AMDGPU::SREG_LIT_0)
+                    .addReg(AMDGPU::SREG_LIT_0)
+                    .addReg(AMDGPU::SREG_LIT_0);
+
+            // ... and terminate wavefront
+            BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::S_ENDPGM));
+          }
+          MI.eraseFromParent();
+          break;
+      }
+    }
+  }
+  return true;
+}
+
+unsigned SILowerControlFlowPass::allocReg() {
+
+  assert(!UnusedRegisters.empty() && "Ran out of registers for predicate stack");
+  unsigned Reg = UnusedRegisters.back();
+  UnusedRegisters.pop_back();
+  return Reg;
+}
+
+void SILowerControlFlowPass::freeReg(unsigned Reg) {
+
+  UnusedRegisters.push_back(Reg);
+}
diff --git a/lib/Target/R600/SILowerLiteralConstants.cpp b/lib/Target/R600/SILowerLiteralConstants.cpp
new file mode 100644
index 0000000000..2d5ab0bea1
--- /dev/null
+++ b/lib/Target/R600/SILowerLiteralConstants.cpp
@@ -0,0 +1,108 @@
+//===-- SILowerLiteralConstants.cpp - Lower intrs using literal constants--===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief This pass performs the following transformation on instructions with
+/// literal constants:
+///
+/// %VGPR0 = V_MOV_IMM_I32 1
+///
+/// becomes:
+///
+/// BUNDLE
+///   * %VGPR = V_MOV_B32_32 SI_LITERAL_CONSTANT
+///   * SI_LOAD_LITERAL 1
+///
+/// The resulting sequence matches exactly how the hardware handles immediate
+/// operands, so this transformation greatly simplifies the code generator.
+///
+/// Only the *_MOV_IMM_* support immediate operands at the moment, but when
+/// support for immediate operands is added to other instructions, they
+/// will be lowered here as well.
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineInstrBundle.h"
+
+using namespace llvm;
+
+namespace {
+
+class SILowerLiteralConstantsPass : public MachineFunctionPass {
+
+private:
+  static char ID;
+  const TargetInstrInfo *TII;
+
+public:
+  SILowerLiteralConstantsPass(TargetMachine &tm) :
+    MachineFunctionPass(ID), TII(tm.getInstrInfo()) { }
+
+  virtual bool runOnMachineFunction(MachineFunction &MF);
+
+  const char *getPassName() const {
+    return "SI Lower literal constants pass";
+  }
+};
+
+} // End anonymous namespace
+
+char SILowerLiteralConstantsPass::ID = 0;
+
+FunctionPass *llvm::createSILowerLiteralConstantsPass(TargetMachine &tm) {
+  return new SILowerLiteralConstantsPass(tm);
+}
+
+bool SILowerLiteralConstantsPass::runOnMachineFunction(MachineFunction &MF) {
+  for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
+                                                  BB != BB_E; ++BB) {
+    MachineBasicBlock &MBB = *BB;
+    for (MachineBasicBlock::iterator I = MBB.begin(), Next = llvm::next(I);
+                               I != MBB.end(); I = Next) {
+      Next = llvm::next(I);
+      MachineInstr &MI = *I;
+      switch (MI.getOpcode()) {
+      default: break;
+      case AMDGPU::S_MOV_IMM_I32:
+      case AMDGPU::S_MOV_IMM_I64:
+      case AMDGPU::V_MOV_IMM_F32:
+      case AMDGPU::V_MOV_IMM_I32: {
+          unsigned MovOpcode;
+          unsigned LoadLiteralOpcode;
+          MachineOperand LiteralOp = MI.getOperand(1);
+          if (AMDGPU::VReg_32RegClass.contains(MI.getOperand(0).getReg())) {
+            MovOpcode = AMDGPU::V_MOV_B32_e32;
+          } else {
+            MovOpcode = AMDGPU::S_MOV_B32;
+          }
+          if (LiteralOp.isImm()) {
+            LoadLiteralOpcode = AMDGPU::SI_LOAD_LITERAL_I32;
+          } else {
+            LoadLiteralOpcode = AMDGPU::SI_LOAD_LITERAL_F32;
+          }
+          MachineInstr *First =
+            BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(MovOpcode),
+                    MI.getOperand(0).getReg())
+                    .addReg(AMDGPU::SI_LITERAL_CONSTANT);
+          MachineInstr *Last =
+            BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(LoadLiteralOpcode))
+                    .addOperand(MI.getOperand(1));
+          Last->setIsInsideBundle();
+          llvm::finalizeBundle(MBB, First, Last);
+          MI.eraseFromParent();
+          break;
+        }
+      }
+    }
+  }
+  return false;
+}
diff --git a/lib/Target/R600/SIMachineFunctionInfo.cpp b/lib/Target/R600/SIMachineFunctionInfo.cpp
new file mode 100644
index 0000000000..7e59b42749
--- /dev/null
+++ b/lib/Target/R600/SIMachineFunctionInfo.cpp
@@ -0,0 +1,20 @@
+//===-- SIMachineFunctionInfo.cpp - SI Machine Function Info -------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+/// \file
+//===----------------------------------------------------------------------===//
+
+
+#include "SIMachineFunctionInfo.h"
+
+using namespace llvm;
+
+SIMachineFunctionInfo::SIMachineFunctionInfo(const MachineFunction &MF)
+  : MachineFunctionInfo(),
+    SPIPSInputAddr(0),
+    ShaderType(0)
+  { }
diff --git a/lib/Target/R600/SIMachineFunctionInfo.h b/lib/Target/R600/SIMachineFunctionInfo.h
new file mode 100644
index 0000000000..47271f5a1e
--- /dev/null
+++ b/lib/Target/R600/SIMachineFunctionInfo.h
@@ -0,0 +1,34 @@
+//===- SIMachineFunctionInfo.h - SIMachineFunctionInfo interface -*- C++ -*-==//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+//
+//===----------------------------------------------------------------------===//
+
+
+#ifndef SIMACHINEFUNCTIONINFO_H_
+#define SIMACHINEFUNCTIONINFO_H_
+
+#include "llvm/CodeGen/MachineFunction.h"
+
+namespace llvm {
+
+/// This class keeps track of the SPI_SP_INPUT_ADDR config register, which
+/// tells the hardware which interpolation parameters to load.
+class SIMachineFunctionInfo : public MachineFunctionInfo {
+public:
+  SIMachineFunctionInfo(const MachineFunction &MF);
+  unsigned SPIPSInputAddr;
+  unsigned ShaderType;
+};
+
+} // End namespace llvm
+
+
+#endif //_SIMACHINEFUNCTIONINFO_H_
diff --git a/lib/Target/R600/SIRegisterInfo.cpp b/lib/Target/R600/SIRegisterInfo.cpp
new file mode 100644
index 0000000000..88275c523f
--- /dev/null
+++ b/lib/Target/R600/SIRegisterInfo.cpp
@@ -0,0 +1,48 @@
+//===-- SIRegisterInfo.cpp - SI Register Information ---------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief SI implementation of the TargetRegisterInfo class.
+//
+//===----------------------------------------------------------------------===//
+
+
+#include "SIRegisterInfo.h"
+#include "AMDGPUTargetMachine.h"
+
+using namespace llvm;
+
+SIRegisterInfo::SIRegisterInfo(AMDGPUTargetMachine &tm,
+    const TargetInstrInfo &tii)
+: AMDGPURegisterInfo(tm, tii),
+  TM(tm),
+  TII(tii)
+  { }
+
+BitVector SIRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
+  BitVector Reserved(getNumRegs());
+  return Reserved;
+}
+
+const TargetRegisterClass *
+SIRegisterInfo::getISARegClass(const TargetRegisterClass * rc) const {
+  switch (rc->getID()) {
+  case AMDGPU::GPRF32RegClassID:
+    return &AMDGPU::VReg_32RegClass;
+  default: return rc;
+  }
+}
+
+const TargetRegisterClass * SIRegisterInfo::getCFGStructurizerRegClass(
+                                                                   MVT VT) const {
+  switch(VT.SimpleTy) {
+    default:
+    case MVT::i32: return &AMDGPU::VReg_32RegClass;
+  }
+}
diff --git a/lib/Target/R600/SIRegisterInfo.h b/lib/Target/R600/SIRegisterInfo.h
new file mode 100644
index 0000000000..40171e4450
--- /dev/null
+++ b/lib/Target/R600/SIRegisterInfo.h
@@ -0,0 +1,47 @@
+//===-- SIRegisterInfo.h - SI Register Info Interface ----------*- C++ -*--===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// \brief Interface definition for SIRegisterInfo
+//
+//===----------------------------------------------------------------------===//
+
+
+#ifndef SIREGISTERINFO_H_
+#define SIREGISTERINFO_H_
+
+#include "AMDGPURegisterInfo.h"
+
+namespace llvm {
+
+class AMDGPUTargetMachine;
+class TargetInstrInfo;
+
+struct SIRegisterInfo : public AMDGPURegisterInfo {
+  AMDGPUTargetMachine &TM;
+  const TargetInstrInfo &TII;
+
+  SIRegisterInfo(AMDGPUTargetMachine &tm, const TargetInstrInfo &tii);
+
+  virtual BitVector getReservedRegs(const MachineFunction &MF) const;
+
+  /// \param RC is an AMDIL reg class.
+  ///
+  /// \returns the SI register class that is equivalent to \p RC.
+  virtual const TargetRegisterClass *
+    getISARegClass(const TargetRegisterClass *RC) const;
+
+  /// \brief get the register class of the specified type to use in the
+  /// CFGStructurizer
+  virtual const TargetRegisterClass * getCFGStructurizerRegClass(MVT VT) const;
+};
+
+} // End namespace llvm
+
+#endif // SIREGISTERINFO_H_
diff --git a/lib/Target/R600/SIRegisterInfo.td b/lib/Target/R600/SIRegisterInfo.td
new file mode 100644
index 0000000000..e52311ab8a
--- /dev/null
+++ b/lib/Target/R600/SIRegisterInfo.td
@@ -0,0 +1,167 @@
+
+let Namespace = "AMDGPU" in {
+  def low : SubRegIndex;
+  def high : SubRegIndex;
+
+  def sub0 : SubRegIndex;
+  def sub1 : SubRegIndex;
+  def sub2 : SubRegIndex;
+  def sub3 : SubRegIndex;
+  def sub4 : SubRegIndex;
+  def sub5 : SubRegIndex;
+  def sub6 : SubRegIndex;
+  def sub7 : SubRegIndex;
+}
+
+class SIReg <string n, bits<16> encoding = 0> : Register<n> {
+  let Namespace = "AMDGPU";
+  let HWEncoding = encoding;
+}
+
+class SI_64 <string n, list<Register> subregs, bits<16> encoding> : RegisterWithSubRegs<n, subregs> {
+  let Namespace = "AMDGPU";
+  let SubRegIndices = [low, high];
+  let HWEncoding = encoding;
+}
+
+class SGPR_32 <bits<16> num, string name> : SIReg<name, num>;
+
+class VGPR_32 <bits<16> num, string name> : SIReg<name, num>;
+
+// Special Registers
+def VCC : SIReg<"VCC", 106>;
+def EXEC_LO : SIReg <"EXEC LO", 126>;
+def EXEC_HI : SIReg <"EXEC HI", 127>;
+def EXEC : SI_64<"EXEC", [EXEC_LO, EXEC_HI], 126>;
+def SCC : SIReg<"SCC", 253>;
+def SREG_LIT_0 : SIReg <"S LIT 0", 128>;
+def SI_LITERAL_CONSTANT : SIReg<"LITERAL CONSTANT", 255>;
+def M0 : SIReg <"M0", 124>;
+
+//Interpolation registers
+def PERSP_SAMPLE_I : SIReg <"PERSP_SAMPLE_I">;
+def PERSP_SAMPLE_J : SIReg <"PERSP_SAMPLE_J">;
+def PERSP_CENTER_I : SIReg <"PERSP_CENTER_I">;
+def PERSP_CENTER_J : SIReg <"PERSP_CENTER_J">;
+def PERSP_CENTROID_I : SIReg <"PERSP_CENTROID_I">;
+def PERSP_CENTROID_J : SIReg <"PERP_CENTROID_J">;
+def PERSP_I_W : SIReg <"PERSP_I_W">;
+def PERSP_J_W : SIReg <"PERSP_J_W">;
+def PERSP_1_W : SIReg <"PERSP_1_W">;
+def LINEAR_SAMPLE_I : SIReg <"LINEAR_SAMPLE_I">;
+def LINEAR_SAMPLE_J : SIReg <"LINEAR_SAMPLE_J">;
+def LINEAR_CENTER_I : SIReg <"LINEAR_CENTER_I">;
+def LINEAR_CENTER_J : SIReg <"LINEAR_CENTER_J">;
+def LINEAR_CENTROID_I : SIReg <"LINEAR_CENTROID_I">;
+def LINEAR_CENTROID_J : SIReg <"LINEAR_CENTROID_J">;
+def LINE_STIPPLE_TEX_COORD : SIReg <"LINE_STIPPLE_TEX_COORD">;
+def POS_X_FLOAT : SIReg <"POS_X_FLOAT">;
+def POS_Y_FLOAT : SIReg <"POS_Y_FLOAT">;
+def POS_Z_FLOAT : SIReg <"POS_Z_FLOAT">;
+def POS_W_FLOAT : SIReg <"POS_W_FLOAT">;
+def FRONT_FACE : SIReg <"FRONT_FACE">;
+def ANCILLARY : SIReg <"ANCILLARY">;
+def SAMPLE_COVERAGE : SIReg <"SAMPLE_COVERAGE">;
+def POS_FIXED_PT : SIReg <"POS_FIXED_PT">;
+
+// SGPR 32-bit registers
+foreach Index = 0-101 in {
+  def SGPR#Index : SGPR_32 <Index, "SGPR"#Index>;
+}
+
+def SGPR_32 : RegisterClass<"AMDGPU", [f32, i32], 32,
+                            (add (sequence "SGPR%u", 0, 101))>;
+
+// SGPR 64-bit registers
+def SGPR_64 : RegisterTuples<[low, high],
+                             [(add (decimate SGPR_32, 2)),
+                              (add(decimate (rotl SGPR_32, 1), 2))]>;
+
+// SGPR 128-bit registers
+def SGPR_128 : RegisterTuples<[sel_x, sel_y, sel_z, sel_w],
+                              [(add (decimate SGPR_32, 4)),
+                               (add (decimate (rotl SGPR_32, 1), 4)),
+                               (add (decimate (rotl SGPR_32, 2), 4)),
+                               (add (decimate (rotl SGPR_32, 3), 4))]>;
+
+// SGPR 256-bit registers
+def SGPR_256 : RegisterTuples<[sub0, sub1, sub2, sub3, sub4, sub5, sub6, sub7],
+                              [(add (decimate SGPR_32, 8)),
+                               (add (decimate (rotl SGPR_32, 1), 8)),
+                               (add (decimate (rotl SGPR_32, 2), 8)),
+                               (add (decimate (rotl SGPR_32, 3), 8)),
+                               (add (decimate (rotl SGPR_32, 4), 8)),
+                               (add (decimate (rotl SGPR_32, 5), 8)),
+                               (add (decimate (rotl SGPR_32, 6), 8)),
+                               (add (decimate (rotl SGPR_32, 7), 8))]>;
+
+// VGPR 32-bit registers
+foreach Index = 0-255 in {
+  def VGPR#Index : VGPR_32 <Index, "VGPR"#Index>;
+}
+
+def VGPR_32 : RegisterClass<"AMDGPU", [f32, i32], 32,
+                            (add (sequence "VGPR%u", 0, 255))>;
+
+// VGPR 64-bit registers
+def VGPR_64 : RegisterTuples<[low, high],
+                             [(add (decimate VGPR_32, 2)),
+                              (add (decimate (rotl VGPR_32, 1), 2))]>;
+
+// VGPR 128-bit registers
+def VGPR_128 : RegisterTuples<[sel_x, sel_y, sel_z, sel_w],
+                              [(add (decimate VGPR_32, 4)),
+                               (add (decimate (rotl VGPR_32, 1), 4)),
+                               (add (decimate (rotl VGPR_32, 2), 4)),
+                               (add (decimate (rotl VGPR_32, 3), 4))]>;
+
+// Register class for all scalar registers (SGPRs + Special Registers)
+def SReg_32 : RegisterClass<"AMDGPU", [f32, i32], 32,
+    (add SGPR_32,  SREG_LIT_0, M0, EXEC_LO, EXEC_HI)
+>;
+
+def SReg_64 : RegisterClass<"AMDGPU", [i64], 64, (add SGPR_64, VCC, EXEC)>;
+
+def SReg_1 : RegisterClass<"AMDGPU", [i1], 1, (add VCC, SGPR_64, EXEC)>;
+
+def SReg_128 : RegisterClass<"AMDGPU", [v4f32, v4i32], 128, (add SGPR_128)>;
+
+def SReg_256 : RegisterClass<"AMDGPU", [v8i32], 256, (add SGPR_256)>;
+
+// Register class for all vector registers (VGPRs + Interploation Registers)
+def VReg_32 : RegisterClass<"AMDGPU", [f32, i32], 32,
+    (add VGPR_32,
+    PERSP_SAMPLE_I, PERSP_SAMPLE_J,
+    PERSP_CENTER_I, PERSP_CENTER_J,
+    PERSP_CENTROID_I, PERSP_CENTROID_J,
+    PERSP_I_W, PERSP_J_W, PERSP_1_W,
+    LINEAR_SAMPLE_I, LINEAR_SAMPLE_J,
+    LINEAR_CENTER_I, LINEAR_CENTER_J,
+    LINEAR_CENTROID_I, LINEAR_CENTROID_J,
+    LINE_STIPPLE_TEX_COORD,
+    POS_X_FLOAT,
+    POS_Y_FLOAT,
+    POS_Z_FLOAT,
+    POS_W_FLOAT,
+    FRONT_FACE,
+    ANCILLARY,
+    SAMPLE_COVERAGE,
+    POS_FIXED_PT
+    )
+>;
+
+def VReg_64 : RegisterClass<"AMDGPU", [i64], 64, (add VGPR_64)>;
+
+def VReg_128 : RegisterClass<"AMDGPU", [v4f32], 128, (add VGPR_128)>;
+
+// AllReg_* - A set of all scalar and vector registers of a given width.
+def AllReg_32 : RegisterClass<"AMDGPU", [f32, i32], 32, (add VReg_32, SReg_32)>;
+
+def AllReg_64 : RegisterClass<"AMDGPU", [f64, i64], 64, (add SReg_64, VReg_64)>;
+
+// Special register classes for predicates and the M0 register
+def SCCReg : RegisterClass<"AMDGPU", [i1], 1, (add SCC)>;
+def VCCReg : RegisterClass<"AMDGPU", [i1], 1, (add VCC)>;
+def EXECReg : RegisterClass<"AMDGPU", [i1], 1, (add EXEC)>;
+def M0Reg : RegisterClass<"AMDGPU", [i32], 32, (add M0)>;
+
diff --git a/lib/Target/R600/SISchedule.td b/lib/Target/R600/SISchedule.td
new file mode 100644
index 0000000000..28b65b8258
--- /dev/null
+++ b/lib/Target/R600/SISchedule.td
@@ -0,0 +1,15 @@
+//===-- SISchedule.td - SI Scheduling definitons -------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// TODO: This is just a place holder for now.
+//
+//===----------------------------------------------------------------------===//
+
+
+def SI_Itin : ProcessorItineraries <[], [], []>;
diff --git a/lib/Target/R600/TargetInfo/AMDGPUTargetInfo.cpp b/lib/Target/R600/TargetInfo/AMDGPUTargetInfo.cpp
new file mode 100644
index 0000000000..46b1f18c62
--- /dev/null
+++ b/lib/Target/R600/TargetInfo/AMDGPUTargetInfo.cpp
@@ -0,0 +1,26 @@
+//===-- TargetInfo/AMDGPUTargetInfo.cpp - TargetInfo for AMDGPU -----------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "llvm/Support/TargetRegistry.h"
+
+using namespace llvm;
+
+/// \brief The target for the AMDGPU backend
+Target llvm::TheAMDGPUTarget;
+
+/// \brief Extern function to initialize the targets for the AMDGPU backend
+extern "C" void LLVMInitializeR600TargetInfo() {
+  RegisterTarget<Triple::r600, false>
+    R600(TheAMDGPUTarget, "r600", "AMD GPUs HD2XXX-HD6XXX");
+}
diff --git a/lib/Target/R600/TargetInfo/CMakeLists.txt b/lib/Target/R600/TargetInfo/CMakeLists.txt
new file mode 100644
index 0000000000..3d1584eba3
--- /dev/null
+++ b/lib/Target/R600/TargetInfo/CMakeLists.txt
@@ -0,0 +1,7 @@
+include_directories( ${CMAKE_CURRENT_BINARY_DIR}/.. ${CMAKE_CURRENT_SOURCE_DIR}/.. )
+
+add_llvm_library(LLVMR600Info
+  AMDGPUTargetInfo.cpp
+  )
+
+add_dependencies(LLVMR600Info AMDGPUCommonTableGen intrinsics_gen)
diff --git a/lib/Target/R600/TargetInfo/LLVMBuild.txt b/lib/Target/R600/TargetInfo/LLVMBuild.txt
new file mode 100644
index 0000000000..4c6fea4aa0
--- /dev/null
+++ b/lib/Target/R600/TargetInfo/LLVMBuild.txt
@@ -0,0 +1,23 @@
+;===- ./lib/Target/R600/TargetInfo/LLVMBuild.txt --------------*- Conf -*--===;
+;
+;                     The LLVM Compiler Infrastructure
+;
+; This file is distributed under the University of Illinois Open Source
+; License. See LICENSE.TXT for details.
+;
+;===------------------------------------------------------------------------===;
+;
+; This is an LLVMBuild description file for the components in this subdirectory.
+;
+; For more information on the LLVMBuild system, please see:
+;
+;   http://llvm.org/docs/LLVMBuild.html
+;
+;===------------------------------------------------------------------------===;
+
+[component_0]
+type = Library
+name = R600Info
+parent = R600
+required_libraries = MC Support
+add_to_library_groups = R600
diff --git a/lib/Target/R600/TargetInfo/Makefile b/lib/Target/R600/TargetInfo/Makefile
new file mode 100644
index 0000000000..b8ac4e7823
--- /dev/null
+++ b/lib/Target/R600/TargetInfo/Makefile
@@ -0,0 +1,15 @@
+##===- lib/Target/AMDGPU/TargetInfo/Makefile ----------------*- Makefile -*-===##
+#
+#                     The LLVM Compiler Infrastructure
+#
+# This file is distributed under the University of Illinois Open Source
+# License. See LICENSE.TXT for details.
+#
+##===----------------------------------------------------------------------===##
+LEVEL = ../../../..
+LIBRARYNAME = LLVMR600Info
+
+# Hack: we need to include 'main' target directory to grab private headers
+CPPFLAGS = -I$(PROJ_OBJ_DIR)/.. -I$(PROJ_SRC_DIR)/..
+
+include $(LEVEL)/Makefile.common
diff --git a/test/CodeGen/R600/add.v4i32.ll b/test/CodeGen/R600/add.v4i32.ll
new file mode 100644
index 0000000000..ac4a87417b
--- /dev/null
+++ b/test/CodeGen/R600/add.v4i32.ll
@@ -0,0 +1,15 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: ADD_INT T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: ADD_INT T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: ADD_INT T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: ADD_INT T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
+  %b_ptr = getelementptr <4 x i32> addrspace(1)* %in, i32 1
+  %a = load <4 x i32> addrspace(1) * %in
+  %b = load <4 x i32> addrspace(1) * %b_ptr
+  %result = add <4 x i32> %a, %b
+  store <4 x i32> %result, <4 x i32> addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/and.v4i32.ll b/test/CodeGen/R600/and.v4i32.ll
new file mode 100644
index 0000000000..662085e2d6
--- /dev/null
+++ b/test/CodeGen/R600/and.v4i32.ll
@@ -0,0 +1,15 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: AND_INT T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: AND_INT T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: AND_INT T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: AND_INT T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
+  %b_ptr = getelementptr <4 x i32> addrspace(1)* %in, i32 1
+  %a = load <4 x i32> addrspace(1) * %in
+  %b = load <4 x i32> addrspace(1) * %b_ptr
+  %result = and <4 x i32> %a, %b
+  store <4 x i32> %result, <4 x i32> addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/fabs.ll b/test/CodeGen/R600/fabs.ll
new file mode 100644
index 0000000000..0407533eaa
--- /dev/null
+++ b/test/CodeGen/R600/fabs.ll
@@ -0,0 +1,16 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: MOV T{{[0-9]+\.[XYZW], \|T[0-9]+\.[XYZW]\|}}
+
+define void @test() {
+   %r0 = call float @llvm.R600.load.input(i32 0)
+   %r1 = call float @fabs( float %r0)
+   call void @llvm.AMDGPU.store.output(float %r1, i32 0)
+   ret void
+}
+
+declare float @llvm.R600.load.input(i32) readnone
+
+declare void @llvm.AMDGPU.store.output(float, i32)
+
+declare float @fabs(float ) readnone
diff --git a/test/CodeGen/R600/fadd.ll b/test/CodeGen/R600/fadd.ll
new file mode 100644
index 0000000000..d7d1b6572c
--- /dev/null
+++ b/test/CodeGen/R600/fadd.ll
@@ -0,0 +1,16 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+; CHECK: ADD T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test() {
+   %r0 = call float @llvm.R600.load.input(i32 0)
+   %r1 = call float @llvm.R600.load.input(i32 1)
+   %r2 = fadd float %r0, %r1
+   call void @llvm.AMDGPU.store.output(float %r2, i32 0)
+   ret void
+}
+
+declare float @llvm.R600.load.input(i32) readnone
+
+declare void @llvm.AMDGPU.store.output(float, i32)
+
diff --git a/test/CodeGen/R600/fadd.v4f32.ll b/test/CodeGen/R600/fadd.v4f32.ll
new file mode 100644
index 0000000000..85dbfd52cb
--- /dev/null
+++ b/test/CodeGen/R600/fadd.v4f32.ll
@@ -0,0 +1,15 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: ADD T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: ADD T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: ADD T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: ADD T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test(<4 x float> addrspace(1)* %out, <4 x float> addrspace(1)* %in) {
+  %b_ptr = getelementptr <4 x float> addrspace(1)* %in, i32 1
+  %a = load <4 x float> addrspace(1) * %in
+  %b = load <4 x float> addrspace(1) * %b_ptr
+  %result = fadd <4 x float> %a, %b
+  store <4 x float> %result, <4 x float> addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/fcmp-cnd.ll b/test/CodeGen/R600/fcmp-cnd.ll
new file mode 100644
index 0000000000..a94cfb5cf2
--- /dev/null
+++ b/test/CodeGen/R600/fcmp-cnd.ll
@@ -0,0 +1,14 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;Not checking arguments 2 and 3 to CNDE, because they may change between
+;registers and literal.x depending on what the optimizer does.
+;CHECK: CNDE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test(i32 addrspace(1)* %out, float addrspace(1)* %in) {
+entry:
+  %0 = load float addrspace(1)* %in
+  %cmp = fcmp oeq float %0, 0.000000e+00
+  %value = select i1 %cmp, i32 2, i32 3 
+  store i32 %value, i32 addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/fcmp-cnde-int-args.ll b/test/CodeGen/R600/fcmp-cnde-int-args.ll
new file mode 100644
index 0000000000..5c981efa9d
--- /dev/null
+++ b/test/CodeGen/R600/fcmp-cnde-int-args.ll
@@ -0,0 +1,16 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+; This test checks a bug in R600TargetLowering::LowerSELECT_CC where the
+; chance to optimize the fcmp + select instructions to CNDE was missed
+; due to the fact that the operands to fcmp and select had different types
+
+;CHECK: CNDE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], literal.x, 0.0, -1}}
+
+define void @test(i32 addrspace(1)* %out, float addrspace(1)* %in) {
+entry:
+  %0 = load float addrspace(1)* %in
+  %cmp = fcmp oeq float %0, 0.000000e+00
+  %value = select i1 %cmp, i32 -1, i32 0
+  store i32 %value, i32 addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/fcmp.ll b/test/CodeGen/R600/fcmp.ll
new file mode 100644
index 0000000000..1dcd07c0b3
--- /dev/null
+++ b/test/CodeGen/R600/fcmp.ll
@@ -0,0 +1,16 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: SETE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: MOV T{{[0-9]+\.[XYZW], -T[0-9]+\.[XYZW]}}
+;CHECK: FLT_TO_INT T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test(i32 addrspace(1)* %out, float addrspace(1)* %in) {
+entry:
+  %0 = load float addrspace(1)* %in
+  %arrayidx1 = getelementptr inbounds float addrspace(1)* %in, i32 1
+  %1 = load float addrspace(1)* %arrayidx1
+  %cmp = fcmp oeq float %0, %1
+  %sext = sext i1 %cmp to i32
+  store i32 %sext, i32 addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/fdiv.v4f32.ll b/test/CodeGen/R600/fdiv.v4f32.ll
new file mode 100644
index 0000000000..b013fd647c
--- /dev/null
+++ b/test/CodeGen/R600/fdiv.v4f32.ll
@@ -0,0 +1,19 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: RECIP_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: MUL NON-IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: RECIP_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: MUL NON-IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: RECIP_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: MUL NON-IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: RECIP_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: MUL NON-IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test(<4 x float> addrspace(1)* %out, <4 x float> addrspace(1)* %in) {
+  %b_ptr = getelementptr <4 x float> addrspace(1)* %in, i32 1
+  %a = load <4 x float> addrspace(1) * %in
+  %b = load <4 x float> addrspace(1) * %b_ptr
+  %result = fdiv <4 x float> %a, %b
+  store <4 x float> %result, <4 x float> addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/floor.ll b/test/CodeGen/R600/floor.ll
new file mode 100644
index 0000000000..845330f284
--- /dev/null
+++ b/test/CodeGen/R600/floor.ll
@@ -0,0 +1,16 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: FLOOR T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test() {
+   %r0 = call float @llvm.R600.load.input(i32 0)
+   %r1 = call float @floor(float %r0)
+   call void @llvm.AMDGPU.store.output(float %r1, i32 0)
+   ret void
+}
+
+declare float @llvm.R600.load.input(i32) readnone
+
+declare void @llvm.AMDGPU.store.output(float, i32)
+
+declare float @floor(float) readonly
diff --git a/test/CodeGen/R600/fmax.ll b/test/CodeGen/R600/fmax.ll
new file mode 100644
index 0000000000..3708f0b9ee
--- /dev/null
+++ b/test/CodeGen/R600/fmax.ll
@@ -0,0 +1,16 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: MAX T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test() {
+   %r0 = call float @llvm.R600.load.input(i32 0)
+   %r1 = call float @llvm.R600.load.input(i32 1)
+   %r2 = fcmp uge float %r0, %r1
+   %r3 = select i1 %r2, float %r0, float %r1
+   call void @llvm.AMDGPU.store.output(float %r3, i32 0)
+   ret void
+}
+
+declare float @llvm.R600.load.input(i32) readnone
+
+declare void @llvm.AMDGPU.store.output(float, i32)
diff --git a/test/CodeGen/R600/fmin.ll b/test/CodeGen/R600/fmin.ll
new file mode 100644
index 0000000000..19d59ab306
--- /dev/null
+++ b/test/CodeGen/R600/fmin.ll
@@ -0,0 +1,16 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: MIN T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test() {
+   %r0 = call float @llvm.R600.load.input(i32 0)
+   %r1 = call float @llvm.R600.load.input(i32 1)
+   %r2 = fcmp uge float %r0, %r1
+   %r3 = select i1 %r2, float %r1, float %r0
+   call void @llvm.AMDGPU.store.output(float %r3, i32 0)
+   ret void
+}
+
+declare float @llvm.R600.load.input(i32) readnone
+
+declare void @llvm.AMDGPU.store.output(float, i32)
diff --git a/test/CodeGen/R600/fmul.ll b/test/CodeGen/R600/fmul.ll
new file mode 100644
index 0000000000..eb1d523c0b
--- /dev/null
+++ b/test/CodeGen/R600/fmul.ll
@@ -0,0 +1,16 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+; CHECK: MUL_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test() {
+   %r0 = call float @llvm.R600.load.input(i32 0)
+   %r1 = call float @llvm.R600.load.input(i32 1)
+   %r2 = fmul float %r0, %r1
+   call void @llvm.AMDGPU.store.output(float %r2, i32 0)
+   ret void
+}
+
+declare float @llvm.R600.load.input(i32) readnone
+
+declare void @llvm.AMDGPU.store.output(float, i32)
+
diff --git a/test/CodeGen/R600/fmul.v4f32.ll b/test/CodeGen/R600/fmul.v4f32.ll
new file mode 100644
index 0000000000..6d44a0c5c7
--- /dev/null
+++ b/test/CodeGen/R600/fmul.v4f32.ll
@@ -0,0 +1,15 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: MUL_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: MUL_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: MUL_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: MUL_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test(<4 x float> addrspace(1)* %out, <4 x float> addrspace(1)* %in) {
+  %b_ptr = getelementptr <4 x float> addrspace(1)* %in, i32 1
+  %a = load <4 x float> addrspace(1) * %in
+  %b = load <4 x float> addrspace(1) * %b_ptr
+  %result = fmul <4 x float> %a, %b
+  store <4 x float> %result, <4 x float> addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/fsub.ll b/test/CodeGen/R600/fsub.ll
new file mode 100644
index 0000000000..0ec1c376df
--- /dev/null
+++ b/test/CodeGen/R600/fsub.ll
@@ -0,0 +1,17 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+; CHECK: MOV T{{[0-9]+\.[XYZW], -T[0-9]+\.[XYZW]}}
+; CHECK: ADD T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test() {
+   %r0 = call float @llvm.R600.load.input(i32 0)
+   %r1 = call float @llvm.R600.load.input(i32 1)
+   %r2 = fsub float %r0, %r1
+   call void @llvm.AMDGPU.store.output(float %r2, i32 0)
+   ret void
+}
+
+declare float @llvm.R600.load.input(i32) readnone
+
+declare void @llvm.AMDGPU.store.output(float, i32)
+
diff --git a/test/CodeGen/R600/fsub.v4f32.ll b/test/CodeGen/R600/fsub.v4f32.ll
new file mode 100644
index 0000000000..612a57e4b6
--- /dev/null
+++ b/test/CodeGen/R600/fsub.v4f32.ll
@@ -0,0 +1,15 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: ADD T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: ADD T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: ADD T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK: ADD T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test(<4 x float> addrspace(1)* %out, <4 x float> addrspace(1)* %in) {
+  %b_ptr = getelementptr <4 x float> addrspace(1)* %in, i32 1
+  %a = load <4 x float> addrspace(1) * %in
+  %b = load <4 x float> addrspace(1) * %b_ptr
+  %result = fsub <4 x float> %a, %b
+  store <4 x float> %result, <4 x float> addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/i8_to_double_to_float.ll b/test/CodeGen/R600/i8_to_double_to_float.ll
new file mode 100644
index 0000000000..39f33227fa
--- /dev/null
+++ b/test/CodeGen/R600/i8_to_double_to_float.ll
@@ -0,0 +1,11 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: UINT_TO_FLT T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test(float addrspace(1)* %out, i8 addrspace(1)* %in) {
+  %1 = load i8 addrspace(1)* %in
+  %2 = uitofp i8 %1 to double
+  %3 = fptrunc double %2 to float
+  store float %3, float addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/icmp-select-sete-reverse-args.ll b/test/CodeGen/R600/icmp-select-sete-reverse-args.ll
new file mode 100644
index 0000000000..aad44d9edf
--- /dev/null
+++ b/test/CodeGen/R600/icmp-select-sete-reverse-args.ll
@@ -0,0 +1,18 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;Test that a select with reversed True/False values is correctly lowered
+;to a SETNE_INT.  There should only be one SETNE_INT instruction.
+
+;CHECK: SETNE_INT T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK_NOT: SETNE_INT
+
+define void @test(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+entry:
+  %0 = load i32 addrspace(1)* %in
+  %arrayidx1 = getelementptr inbounds i32 addrspace(1)* %in, i32 1
+  %1 = load i32 addrspace(1)* %arrayidx1
+  %cmp = icmp eq i32 %0, %1
+  %value = select i1 %cmp, i32 0, i32 -1
+  store i32 %value, i32 addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/lit.local.cfg b/test/CodeGen/R600/lit.local.cfg
new file mode 100644
index 0000000000..36ee493e59
--- /dev/null
+++ b/test/CodeGen/R600/lit.local.cfg
@@ -0,0 +1,13 @@
+config.suffixes = ['.ll', '.c', '.cpp']
+
+def getRoot(config):
+    if not config.parent:
+        return config
+    return getRoot(config.parent)
+
+root = getRoot(config)
+
+targets = set(root.targets_to_build.split())
+if not 'R600' in targets:
+    config.unsupported = True
+
diff --git a/test/CodeGen/R600/literals.ll b/test/CodeGen/R600/literals.ll
new file mode 100644
index 0000000000..4c731b25ec
--- /dev/null
+++ b/test/CodeGen/R600/literals.ll
@@ -0,0 +1,30 @@
+; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+; Test using an integer literal constant.
+; Generated ASM should be:
+; ADD_INT REG literal.x, 5
+; or
+; ADD_INT literal.x REG, 5
+
+; CHECK: ADD_INT {{[A-Z0-9,. ]*}}literal.x,{{[A-Z0-9,. ]*}} 5
+define void @i32_literal(i32 addrspace(1)* %out, i32 %in) {
+entry:
+  %0 = add i32 5, %in
+  store i32 %0, i32 addrspace(1)* %out
+  ret void
+}
+
+; Test using a float literal constant.
+; Generated ASM should be:
+; ADD REG literal.x, 5.0
+; or
+; ADD literal.x REG, 5.0
+
+; CHECK: ADD {{[A-Z0-9,. ]*}}literal.x,{{[A-Z0-9,. ]*}} {{[0-9]+}}(5.0
+define void @float_literal(float addrspace(1)* %out, float %in) {
+entry:
+  %0 = fadd float 5.0, %in
+  store float %0, float addrspace(1)* %out
+  ret void
+}
+
diff --git a/test/CodeGen/R600/llvm.AMDGPU.mul.ll b/test/CodeGen/R600/llvm.AMDGPU.mul.ll
new file mode 100644
index 0000000000..693eb27457
--- /dev/null
+++ b/test/CodeGen/R600/llvm.AMDGPU.mul.ll
@@ -0,0 +1,17 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: MUL NON-IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test() {
+   %r0 = call float @llvm.R600.load.input(i32 0)
+   %r1 = call float @llvm.R600.load.input(i32 1)
+   %r2 = call float @llvm.AMDGPU.mul( float %r0, float %r1)
+   call void @llvm.AMDGPU.store.output(float %r2, i32 0)
+   ret void
+}
+
+declare float @llvm.R600.load.input(i32) readnone
+
+declare void @llvm.AMDGPU.store.output(float, i32)
+
+declare float @llvm.AMDGPU.mul(float ,float ) readnone
diff --git a/test/CodeGen/R600/llvm.AMDGPU.trunc.ll b/test/CodeGen/R600/llvm.AMDGPU.trunc.ll
new file mode 100644
index 0000000000..fac957f7ee
--- /dev/null
+++ b/test/CodeGen/R600/llvm.AMDGPU.trunc.ll
@@ -0,0 +1,16 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: TRUNC T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test() {
+   %r0 = call float @llvm.R600.load.input(i32 0)
+   %r1 = call float @llvm.AMDGPU.trunc( float %r0)
+   call void @llvm.AMDGPU.store.output(float %r1, i32 0)
+   ret void
+}
+
+declare float @llvm.R600.load.input(i32) readnone
+
+declare void @llvm.AMDGPU.store.output(float, i32)
+
+declare float @llvm.AMDGPU.trunc(float ) readnone
diff --git a/test/CodeGen/R600/llvm.cos.ll b/test/CodeGen/R600/llvm.cos.ll
new file mode 100644
index 0000000000..dc120bfb00
--- /dev/null
+++ b/test/CodeGen/R600/llvm.cos.ll
@@ -0,0 +1,16 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: COS T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test() {
+   %r0 = call float @llvm.R600.load.input(i32 0)
+   %r1 = call float @llvm.cos.f32(float %r0)
+   call void @llvm.AMDGPU.store.output(float %r1, i32 0)
+   ret void
+}
+
+declare float @llvm.cos.f32(float) readnone
+
+declare float @llvm.R600.load.input(i32) readnone
+
+declare void @llvm.AMDGPU.store.output(float, i32)
diff --git a/test/CodeGen/R600/llvm.pow.ll b/test/CodeGen/R600/llvm.pow.ll
new file mode 100644
index 0000000000..0ae9172579
--- /dev/null
+++ b/test/CodeGen/R600/llvm.pow.ll
@@ -0,0 +1,19 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: LOG_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK-NEXT: MUL_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+;CHECK-NEXT: EXP_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test() {
+   %r0 = call float @llvm.R600.load.input(i32 0)
+   %r1 = call float @llvm.R600.load.input(i32 1)
+   %r2 = call float @llvm.pow.f32( float %r0, float %r1)
+   call void @llvm.AMDGPU.store.output(float %r2, i32 0)
+   ret void
+}
+
+declare float @llvm.R600.load.input(i32) readnone
+
+declare void @llvm.AMDGPU.store.output(float, i32)
+
+declare float @llvm.pow.f32(float ,float ) readonly
diff --git a/test/CodeGen/R600/llvm.sin.ll b/test/CodeGen/R600/llvm.sin.ll
new file mode 100644
index 0000000000..5cd6998c93
--- /dev/null
+++ b/test/CodeGen/R600/llvm.sin.ll
@@ -0,0 +1,16 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: SIN T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test() {
+   %r0 = call float @llvm.R600.load.input(i32 0)
+   %r1 = call float @llvm.sin.f32( float %r0)
+   call void @llvm.AMDGPU.store.output(float %r1, i32 0)
+   ret void
+}
+
+declare float @llvm.sin.f32(float) readnone
+
+declare float @llvm.R600.load.input(i32) readnone
+
+declare void @llvm.AMDGPU.store.output(float, i32)
diff --git a/test/CodeGen/R600/load.constant_addrspace.f32.ll b/test/CodeGen/R600/load.constant_addrspace.f32.ll
new file mode 100644
index 0000000000..93627283bb
--- /dev/null
+++ b/test/CodeGen/R600/load.constant_addrspace.f32.ll
@@ -0,0 +1,9 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: VTX_READ_32 T{{[0-9]+\.X, T[0-9]+\.X}}
+
+define void @test(float addrspace(1)* %out, float addrspace(2)* %in) {
+  %1 = load float addrspace(2)* %in
+  store float %1, float addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/load.i8.ll b/test/CodeGen/R600/load.i8.ll
new file mode 100644
index 0000000000..b070dcd520
--- /dev/null
+++ b/test/CodeGen/R600/load.i8.ll
@@ -0,0 +1,10 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: VTX_READ_8 T{{[0-9]+\.X, T[0-9]+\.X}}
+
+define void @test(i32 addrspace(1)* %out, i8 addrspace(1)* %in) {
+  %1 = load i8 addrspace(1)* %in
+  %2 = zext i8 %1 to i32
+  store i32 %2, i32 addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/reciprocal.ll b/test/CodeGen/R600/reciprocal.ll
new file mode 100644
index 0000000000..6838c1ae36
--- /dev/null
+++ b/test/CodeGen/R600/reciprocal.ll
@@ -0,0 +1,16 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: RECIP_IEEE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test() {
+   %r0 = call float @llvm.R600.load.input(i32 0)
+   %r1 = fdiv float 1.0, %r0
+   call void @llvm.AMDGPU.store.output(float %r1, i32 0)
+   ret void
+}
+
+declare float @llvm.R600.load.input(i32) readnone
+
+declare void @llvm.AMDGPU.store.output(float, i32)
+
+declare float @llvm.AMDGPU.rcp(float ) readnone
diff --git a/test/CodeGen/R600/sdiv.ll b/test/CodeGen/R600/sdiv.ll
new file mode 100644
index 0000000000..3556facfba
--- /dev/null
+++ b/test/CodeGen/R600/sdiv.ll
@@ -0,0 +1,21 @@
+; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+; The code generated by sdiv is long and complex and may frequently change.
+; The goal of this test is to make sure the ISel doesn't fail.
+;
+; This program was previously failing to compile when one of the selectcc
+; opcodes generated by the sdiv lowering was being legalized and optimized to:
+; selectcc Remainder -1, 0, -1, SETGT
+; This was fixed by adding an additional pattern in R600Instructions.td to
+; match this pattern with a CNDGE_INT.
+
+; CHECK: RETURN
+
+define void @test(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+  %den_ptr = getelementptr i32 addrspace(1)* %in, i32 1
+  %num = load i32 addrspace(1) * %in
+  %den = load i32 addrspace(1) * %den_ptr
+  %result = sdiv i32 %num, %den
+  store i32 %result, i32 addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/selectcc-icmp-select-float.ll b/test/CodeGen/R600/selectcc-icmp-select-float.ll
new file mode 100644
index 0000000000..f65a30086e
--- /dev/null
+++ b/test/CodeGen/R600/selectcc-icmp-select-float.ll
@@ -0,0 +1,15 @@
+; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+; Note additional optimizations may cause this SGT to be replaced with a
+; CND* instruction.
+; CHECK: SGT_INT T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], literal.x, -1}}
+; Test a selectcc with i32 LHS/RHS and float True/False
+
+define void @test(float addrspace(1)* %out, i32 addrspace(1)* %in) {
+entry:
+  %0 = load i32 addrspace(1)* %in
+  %1 = icmp sge i32 %0, 0
+  %2 = select i1 %1, float 1.0, float 0.0
+  store float %2, float addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/selectcc_cnde.ll b/test/CodeGen/R600/selectcc_cnde.ll
new file mode 100644
index 0000000000..f0a0f512ba
--- /dev/null
+++ b/test/CodeGen/R600/selectcc_cnde.ll
@@ -0,0 +1,11 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK-NOT: SETE
+;CHECK: CNDE T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], 1.0, literal.x, [-0-9]+\(2.0}}
+define void @test(float addrspace(1)* %out, float addrspace(1)* %in) {
+  %1 = load float addrspace(1)* %in
+  %2 = fcmp oeq float %1, 0.0
+  %3 = select i1 %2, float 1.0, float 2.0
+  store float %3, float addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/selectcc_cnde_int.ll b/test/CodeGen/R600/selectcc_cnde_int.ll
new file mode 100644
index 0000000000..b38078e26d
--- /dev/null
+++ b/test/CodeGen/R600/selectcc_cnde_int.ll
@@ -0,0 +1,11 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK-NOT: SETE_INT
+;CHECK: CNDE_INT T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], 1, literal.x, 2}}
+define void @test(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
+  %1 = load i32 addrspace(1)* %in
+  %2 = icmp eq i32 %1, 0
+  %3 = select i1 %2, i32 1, i32 2
+  store i32 %3, i32 addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/setcc.v4i32.ll b/test/CodeGen/R600/setcc.v4i32.ll
new file mode 100644
index 0000000000..0752f2e63d
--- /dev/null
+++ b/test/CodeGen/R600/setcc.v4i32.ll
@@ -0,0 +1,12 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+;CHECK: SETE_INT T{{[0-9]+\.[XYZW], T[0-9]+\.[XYZW], T[0-9]+\.[XYZW]}}
+
+define void @test(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
+  %b_ptr = getelementptr <4 x i32> addrspace(1)* %in, i32 1
+  %a = load <4 x i32> addrspace(1) * %in
+  %b = load <4 x i32> addrspace(1) * %b_ptr
+  %result = icmp eq <4 x i32> %a, %b
+  %sext = sext <4 x i1> %result to <4 x i32>
+  store <4 x i32> %sext, <4 x i32> addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/short-args.ll b/test/CodeGen/R600/short-args.ll
new file mode 100644
index 0000000000..107025045c
--- /dev/null
+++ b/test/CodeGen/R600/short-args.ll
@@ -0,0 +1,37 @@
+; RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+; CHECK: VTX_READ_8 T{{[0-9]+\.X, T[0-9]+\.X}}
+
+define void @i8_arg(i32 addrspace(1)* nocapture %out, i8 %in) nounwind {
+entry:
+  %0 = zext i8 %in to i32
+  store i32 %0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; CHECK: VTX_READ_8 T{{[0-9]+\.X, T[0-9]+\.X}}
+
+define void @i8_zext_arg(i32 addrspace(1)* nocapture %out, i8 zeroext %in) nounwind {
+entry:
+  %0 = zext i8 %in to i32
+  store i32 %0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; CHECK: VTX_READ_16 T{{[0-9]+\.X, T[0-9]+\.X}}
+
+define void @i16_arg(i32 addrspace(1)* nocapture %out, i16 %in) nounwind {
+entry:
+  %0 = zext i16 %in to i32
+  store i32 %0, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; CHECK: VTX_READ_16 T{{[0-9]+\.X, T[0-9]+\.X}}
+
+define void @i16_zext_arg(i32 addrspace(1)* nocapture %out, i16 zeroext %in) nounwind {
+entry:
+  %0 = zext i16 %in to i32
+  store i32 %0, i32 addrspace(1)* %out, align 4
+  ret void
+}
diff --git a/test/CodeGen/R600/store.v4f32.ll b/test/CodeGen/R600/store.v4f32.ll
new file mode 100644
index 0000000000..8b0d244459
--- /dev/null
+++ b/test/CodeGen/R600/store.v4f32.ll
@@ -0,0 +1,9 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: RAT_WRITE_CACHELESS_128 T{{[0-9]+\.XYZW, T[0-9]+\.X}}, 1
+
+define void @test(<4 x float> addrspace(1)* %out, <4 x float> addrspace(1)* %in) {
+  %1 = load <4 x float> addrspace(1) * %in
+  store <4 x float> %1, <4 x float> addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/store.v4i32.ll b/test/CodeGen/R600/store.v4i32.ll
new file mode 100644
index 0000000000..a659815dde
--- /dev/null
+++ b/test/CodeGen/R600/store.v4i32.ll
@@ -0,0 +1,9 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;CHECK: RAT_WRITE_CACHELESS_128 T{{[0-9]+\.XYZW, T[0-9]+\.X}}, 1
+
+define void @test(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
+  %1 = load <4 x i32> addrspace(1) * %in
+  store <4 x i32> %1, <4 x i32> addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/udiv.v4i32.ll b/test/CodeGen/R600/udiv.v4i32.ll
new file mode 100644
index 0000000000..47657a6be7
--- /dev/null
+++ b/test/CodeGen/R600/udiv.v4i32.ll
@@ -0,0 +1,15 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;The code generated by udiv is long and complex and may frequently change.
+;The goal of this test is to make sure the ISel doesn't fail when it gets
+;a v4i32 udiv
+;CHECK: RETURN
+
+define void @test(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
+  %b_ptr = getelementptr <4 x i32> addrspace(1)* %in, i32 1
+  %a = load <4 x i32> addrspace(1) * %in
+  %b = load <4 x i32> addrspace(1) * %b_ptr
+  %result = udiv <4 x i32> %a, %b
+  store <4 x i32> %result, <4 x i32> addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/R600/urem.v4i32.ll b/test/CodeGen/R600/urem.v4i32.ll
new file mode 100644
index 0000000000..2e7388caa6
--- /dev/null
+++ b/test/CodeGen/R600/urem.v4i32.ll
@@ -0,0 +1,15 @@
+;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
+
+;The code generated by urem is long and complex and may frequently change.
+;The goal of this test is to make sure the ISel doesn't fail when it gets
+;a v4i32 urem
+;CHECK: RETURN
+
+define void @test(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
+  %b_ptr = getelementptr <4 x i32> addrspace(1)* %in, i32 1
+  %a = load <4 x i32> addrspace(1) * %in
+  %b = load <4 x i32> addrspace(1) * %b_ptr
+  %result = urem <4 x i32> %a, %b
+  store <4 x i32> %result, <4 x i32> addrspace(1)* %out
+  ret void
+}
diff --git a/test/CodeGen/SI/sanity.ll b/test/CodeGen/SI/sanity.ll
new file mode 100644
index 0000000000..62cdcf5eca
--- /dev/null
+++ b/test/CodeGen/SI/sanity.ll
@@ -0,0 +1,37 @@
+;RUN: llc < %s -march=r600 -mcpu=SI | FileCheck %s
+
+; CHECK: S_ENDPGM
+
+define void @main() {
+main_body:
+  call void @llvm.AMDGPU.shader.type(i32 1)
+  %0 = load <4 x i32> addrspace(2)* addrspace(8)* inttoptr (i32 6 to <4 x i32> addrspace(2)* addrspace(8)*)
+  %1 = getelementptr <4 x i32> addrspace(2)* %0, i32 0
+  %2 = load <4 x i32> addrspace(2)* %1
+  %3 = call i32 @llvm.SI.vs.load.buffer.index()
+  %4 = call <4 x float> @llvm.SI.vs.load.input(<4 x i32> %2, i32 0, i32 %3)
+  %5 = extractelement <4 x float> %4, i32 0
+  %6 = extractelement <4 x float> %4, i32 1
+  %7 = extractelement <4 x float> %4, i32 2
+  %8 = extractelement <4 x float> %4, i32 3
+  %9 = load <4 x i32> addrspace(2)* addrspace(8)* inttoptr (i32 6 to <4 x i32> addrspace(2)* addrspace(8)*)
+  %10 = getelementptr <4 x i32> addrspace(2)* %9, i32 1
+  %11 = load <4 x i32> addrspace(2)* %10
+  %12 = call i32 @llvm.SI.vs.load.buffer.index()
+  %13 = call <4 x float> @llvm.SI.vs.load.input(<4 x i32> %11, i32 0, i32 %12)
+  %14 = extractelement <4 x float> %13, i32 0
+  %15 = extractelement <4 x float> %13, i32 1
+  %16 = extractelement <4 x float> %13, i32 2
+  %17 = extractelement <4 x float> %13, i32 3
+  call void @llvm.SI.export(i32 15, i32 0, i32 0, i32 32, i32 0, float %14, float %15, float %16, float %17)
+  call void @llvm.SI.export(i32 15, i32 0, i32 1, i32 12, i32 0, float %5, float %6, float %7, float %8)
+  ret void
+}
+
+declare void @llvm.AMDGPU.shader.type(i32)
+
+declare i32 @llvm.SI.vs.load.buffer.index() readnone
+
+declare <4 x float> @llvm.SI.vs.load.input(<4 x i32>, i32, i32)
+
+declare void @llvm.SI.export(i32, i32, i32, i32, i32, float, float, float, float)