R600: Support for indirect addressing v4

Only implemented for R600 so far.  SI is missing implementations of a
few callbacks used by the Indirect Addressing pass and needs code to
handle frame indices.

At the moment R600 only supports array sizes of 16 dwords or less.
Register packing of vector types is currently disabled, which means that a
vec4 is stored in T0_X, T1_X, T2_X, T3_X, rather than T0_XYZW. In order
to correctly pack registers in all cases, we will need to implement an
analysis pass for R600 that determines the correct vector width for each
array.

v2:
  - Add support for i8 zext load from stack.
  - Coding style fixes

v3:
  - Don't reserve registers for indirect addressing when it isn't
    being used.
  - Fix bug caused by LLVM limiting the number of SubRegIndex
    declarations.

v4:
  - Fix 64-bit defines

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@174525 91177308-0d34-0410-b5e6-96231b3b80d8

30 files changed, 1124 insertions, 75 deletions

diff --git a/lib/Target/R600/AMDGPU.h b/lib/Target/R600/AMDGPU.h
index 1aa607f57e..bac01a3f42 100644
--- a/lib/Target/R600/AMDGPU.h
+++ b/lib/Target/R600/AMDGPU.h
@@ -36,6 +36,7 @@ FunctionPass *createSIInsertWaits(TargetMachine &tm);
 // Passes common to R600 and SI
 Pass *createAMDGPUStructurizeCFGPass();
 FunctionPass *createAMDGPUConvertToISAPass(TargetMachine &tm);
+FunctionPass* createAMDGPUIndirectAddressingPass(TargetMachine &tm);
 
 } // End namespace llvm
 
diff --git a/lib/Target/R600/AMDGPUFrameLowering.cpp b/lib/Target/R600/AMDGPUFrameLowering.cpp
new file mode 100644
index 0000000000..815d6f71c3
--- /dev/null
+++ b/lib/Target/R600/AMDGPUFrameLowering.cpp
@@ -0,0 +1,122 @@
+//===----------------------- AMDGPUFrameLowering.cpp ----------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//==-----------------------------------------------------------------------===//
+//
+// Interface to describe a layout of a stack frame on a AMDIL target machine
+//
+//===----------------------------------------------------------------------===//
+#include "AMDGPUFrameLowering.h"
+#include "AMDGPURegisterInfo.h"
+#include "R600MachineFunctionInfo.h"
+#include "llvm/CodeGen/MachineFrameInfo.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/IR/Instructions.h"
+
+using namespace llvm;
+AMDGPUFrameLowering::AMDGPUFrameLowering(StackDirection D, unsigned StackAl,
+    int LAO, unsigned TransAl)
+  : TargetFrameLowering(D, StackAl, LAO, TransAl) { }
+
+AMDGPUFrameLowering::~AMDGPUFrameLowering() { }
+
+unsigned AMDGPUFrameLowering::getStackWidth(const MachineFunction &MF) const {
+
+  // XXX: Hardcoding to 1 for now.
+  //
+  // I think the StackWidth should stored as metadata associated with the
+  // MachineFunction.  This metadata can either be added by a frontend, or
+  // calculated by a R600 specific LLVM IR pass.
+  //
+  // The StackWidth determines how stack objects are laid out in memory.
+  // For a vector stack variable, like: int4 stack[2], the data will be stored
+  // in the following ways depending on the StackWidth.
+  //
+  // StackWidth = 1:
+  //
+  // T0.X = stack[0].x
+  // T1.X = stack[0].y
+  // T2.X = stack[0].z
+  // T3.X = stack[0].w
+  // T4.X = stack[1].x
+  // T5.X = stack[1].y
+  // T6.X = stack[1].z
+  // T7.X = stack[1].w
+  //
+  // StackWidth = 2:
+  //
+  // T0.X = stack[0].x
+  // T0.Y = stack[0].y
+  // T1.X = stack[0].z
+  // T1.Y = stack[0].w
+  // T2.X = stack[1].x
+  // T2.Y = stack[1].y
+  // T3.X = stack[1].z
+  // T3.Y = stack[1].w
+  // 
+  // StackWidth = 4:
+  // T0.X = stack[0].x
+  // T0.Y = stack[0].y
+  // T0.Z = stack[0].z
+  // T0.W = stack[0].w
+  // T1.X = stack[1].x
+  // T1.Y = stack[1].y
+  // T1.Z = stack[1].z
+  // T1.W = stack[1].w
+  return 1;
+}
+
+/// \returns The number of registers allocated for \p FI.
+int AMDGPUFrameLowering::getFrameIndexOffset(const MachineFunction &MF,
+                                         int FI) const {
+  const MachineFrameInfo *MFI = MF.getFrameInfo();
+  unsigned Offset = 0;
+  int UpperBound = FI == -1 ? MFI->getNumObjects() : FI;
+
+  for (int i = MFI->getObjectIndexBegin(); i < UpperBound; ++i) {
+    const AllocaInst *Alloca = MFI->getObjectAllocation(i);
+    unsigned ArrayElements;
+    const Type *AllocaType = Alloca->getAllocatedType();
+    const Type *ElementType;
+
+    if (AllocaType->isArrayTy()) {
+      ArrayElements = AllocaType->getArrayNumElements();
+      ElementType = AllocaType->getArrayElementType();
+    } else {
+      ArrayElements = 1;
+      ElementType = AllocaType;
+    }
+
+    unsigned VectorElements;
+    if (ElementType->isVectorTy()) {
+      VectorElements = ElementType->getVectorNumElements();
+    } else {
+      VectorElements = 1;
+    }
+
+    Offset += (VectorElements / getStackWidth(MF)) * ArrayElements;
+  }
+  return Offset;
+}
+
+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/AMDGPUFrameLowering.h
index 51337c3dd2..cf5742ee09 100644
--- a/lib/Target/R600/AMDILFrameLowering.h
+++ b/lib/Target/R600/AMDGPUFrameLowering.h
@@ -1,4 +1,4 @@
-//===--------------------- AMDILFrameLowering.h -----------------*- C++ -*-===//
+//===--------------------- AMDGPUFrameLowering.h ----------------*- C++ -*-===//
 //
 //                     The LLVM Compiler Infrastructure
 //
@@ -30,6 +30,10 @@ public:
   AMDGPUFrameLowering(StackDirection D, unsigned StackAl, int LAO,
                       unsigned TransAl = 1);
   virtual ~AMDGPUFrameLowering();
+
+  /// \returns The number of 32-bit sub-registers that are used when storing
+  /// values to the stack.
+  virtual unsigned getStackWidth(const MachineFunction &MF) const;
   virtual int getFrameIndexOffset(const MachineFunction &MF, int FI) const;
   virtual const SpillSlot *getCalleeSavedSpillSlots(unsigned &NumEntries) const;
   virtual void emitPrologue(MachineFunction &MF) const;
diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
index f3a047a135..d0d23d692a 100644
--- a/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -412,5 +412,7 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
   NODE_NAME_CASE(URECIP)
   NODE_NAME_CASE(EXPORT)
   NODE_NAME_CASE(CONST_ADDRESS)
+  NODE_NAME_CASE(REGISTER_LOAD)
+  NODE_NAME_CASE(REGISTER_STORE)
   }
 }
diff --git a/lib/Target/R600/AMDGPUISelLowering.h b/lib/Target/R600/AMDGPUISelLowering.h
index 0584d39682..927ed09b46 100644
--- a/lib/Target/R600/AMDGPUISelLowering.h
+++ b/lib/Target/R600/AMDGPUISelLowering.h
@@ -122,6 +122,8 @@ enum {
   URECIP,
   EXPORT,
   CONST_ADDRESS,
+  REGISTER_LOAD,
+  REGISTER_STORE,
   LAST_AMDGPU_ISD_NUMBER
 };
 
diff --git a/lib/Target/R600/AMDGPUIndirectAddressing.cpp b/lib/Target/R600/AMDGPUIndirectAddressing.cpp
new file mode 100644
index 0000000000..56aaf23cae
--- /dev/null
+++ b/lib/Target/R600/AMDGPUIndirectAddressing.cpp
@@ -0,0 +1,326 @@
+//===-- AMDGPUIndirectAddressing.cpp - Indirect Adressing Support ---------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+///
+/// Instructions can use indirect addressing to index the register file as if it
+/// were memory.  This pass lowers RegisterLoad and RegisterStore instructions
+/// to either a COPY or a MOV that uses indirect addressing.
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "R600InstrInfo.h"
+#include "R600MachineFunctionInfo.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/Support/Debug.h"
+
+using namespace llvm;
+
+namespace {
+
+class AMDGPUIndirectAddressingPass : public MachineFunctionPass {
+
+private:
+  static char ID;
+  const AMDGPUInstrInfo *TII;
+
+  bool regHasExplicitDef(MachineRegisterInfo &MRI, unsigned Reg) const;
+
+public:
+  AMDGPUIndirectAddressingPass(TargetMachine &tm) :
+    MachineFunctionPass(ID),
+    TII(static_cast<const AMDGPUInstrInfo*>(tm.getInstrInfo()))
+    { }
+
+  virtual bool runOnMachineFunction(MachineFunction &MF);
+
+  const char *getPassName() const { return "R600 Handle indirect addressing"; }
+
+};
+
+} // End anonymous namespace
+
+char AMDGPUIndirectAddressingPass::ID = 0;
+
+FunctionPass *llvm::createAMDGPUIndirectAddressingPass(TargetMachine &tm) {
+  return new AMDGPUIndirectAddressingPass(tm);
+}
+
+bool AMDGPUIndirectAddressingPass::runOnMachineFunction(MachineFunction &MF) {
+  MachineRegisterInfo &MRI = MF.getRegInfo();
+
+  int IndirectBegin = TII->getIndirectIndexBegin(MF);
+  int IndirectEnd = TII->getIndirectIndexEnd(MF);
+
+  if (IndirectBegin == -1) {
+    // No indirect addressing, we can skip this pass
+    assert(IndirectEnd == -1);
+    return false;
+  }
+
+  // The map keeps track of the indirect address that is represented by
+  // each virtual register. The key is the register and the value is the
+  // indirect address it uses.
+  std::map<unsigned, unsigned> RegisterAddressMap;
+
+  // First pass - Lower all of the RegisterStore instructions and track which
+  // registers are live.
+  for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
+                                                      BB != BB_E; ++BB) {
+    // This map keeps track of the current live indirect registers.
+    // The key is the address and the value is the register
+    std::map<unsigned, unsigned> LiveAddressRegisterMap;
+    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;
+
+      if (!TII->isRegisterStore(MI)) {
+        continue;
+      }
+
+      // Lower RegisterStore
+
+      unsigned RegIndex = MI.getOperand(2).getImm();
+      unsigned Channel = MI.getOperand(3).getImm();
+      unsigned Address = TII->calculateIndirectAddress(RegIndex, Channel);
+      const TargetRegisterClass *IndirectStoreRegClass =
+                   TII->getIndirectAddrStoreRegClass(MI.getOperand(0).getReg());
+
+      if (MI.getOperand(1).getReg() == AMDGPU::INDIRECT_BASE_ADDR) {
+        // Direct register access.
+        unsigned DstReg = MRI.createVirtualRegister(IndirectStoreRegClass);
+
+        BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::COPY), DstReg)
+                .addOperand(MI.getOperand(0));
+
+        RegisterAddressMap[DstReg] = Address;
+        LiveAddressRegisterMap[Address] = DstReg;
+      } else {
+        // Indirect register access.
+        MachineInstrBuilder MOV = TII->buildIndirectWrite(BB, I,
+                                           MI.getOperand(0).getReg(), // Value
+                                           Address,
+                                           MI.getOperand(1).getReg()); // Offset
+        for (int i = IndirectBegin; i <= IndirectEnd; ++i) {
+          unsigned Addr = TII->calculateIndirectAddress(i, Channel);
+          unsigned DstReg = MRI.createVirtualRegister(IndirectStoreRegClass);
+          MOV.addReg(DstReg, RegState::Define | RegState::Implicit);
+          RegisterAddressMap[DstReg] = Addr;
+          LiveAddressRegisterMap[Addr] = DstReg;
+        }
+      }
+      MI.eraseFromParent();
+    }
+
+    // Update the live-ins of the succesor blocks
+    for (MachineBasicBlock::succ_iterator Succ = MBB.succ_begin(),
+                                          SuccEnd = MBB.succ_end();
+                                          SuccEnd != Succ; ++Succ) {
+      std::map<unsigned, unsigned>::const_iterator Key, KeyEnd;
+      for (Key = LiveAddressRegisterMap.begin(),
+           KeyEnd = LiveAddressRegisterMap.end(); KeyEnd != Key; ++Key) {
+        (*Succ)->addLiveIn(Key->second);
+      }
+    }
+  }
+
+  // Second pass - Lower the RegisterLoad instructions
+  for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
+                                                      BB != BB_E; ++BB) {
+    // Key is the address and the value is the register
+    std::map<unsigned, unsigned> LiveAddressRegisterMap;
+    MachineBasicBlock &MBB = *BB;
+
+    MachineBasicBlock::livein_iterator LI = MBB.livein_begin();
+    while (LI != MBB.livein_end()) {
+      std::vector<unsigned> PhiRegisters;
+
+      // Make sure this live in is used for indirect addressing
+      if (RegisterAddressMap.find(*LI) == RegisterAddressMap.end()) {
+        ++LI;
+        continue;
+      }
+
+      unsigned Address = RegisterAddressMap[*LI];
+      LiveAddressRegisterMap[Address] = *LI;
+      PhiRegisters.push_back(*LI);
+
+      // Check if there are other live in registers which map to the same
+      // indirect address.
+      for (MachineBasicBlock::livein_iterator LJ = llvm::next(LI),
+                                              LE = MBB.livein_end();
+                                              LJ != LE; ++LJ) {
+        unsigned Reg = *LJ;
+        if (RegisterAddressMap.find(Reg) == RegisterAddressMap.end()) {
+          continue;
+        }
+
+        if (RegisterAddressMap[Reg] == Address) {
+          if (!regHasExplicitDef(MRI, Reg)) {
+            continue;
+          }
+          PhiRegisters.push_back(Reg);
+        }
+      }
+
+      if (PhiRegisters.size() == 1) {
+        // We don't need to insert a Phi instruction, so we can just add the
+        // registers to the live list for the block.
+        LiveAddressRegisterMap[Address] = *LI;
+        MBB.removeLiveIn(*LI);
+      } else {
+        // We need to insert a PHI, because we have the same address being
+        // written in multiple predecessor blocks.
+        const TargetRegisterClass *PhiDstClass =
+                   TII->getIndirectAddrStoreRegClass(*(PhiRegisters.begin()));
+        unsigned PhiDstReg = MRI.createVirtualRegister(PhiDstClass);
+        MachineInstrBuilder Phi = BuildMI(MBB, MBB.begin(),
+                                          MBB.findDebugLoc(MBB.begin()),
+                                          TII->get(AMDGPU::PHI), PhiDstReg);
+
+        for (std::vector<unsigned>::const_iterator RI = PhiRegisters.begin(),
+                                                   RE = PhiRegisters.end();
+                                                   RI != RE; ++RI) {
+          unsigned Reg = *RI;
+          MachineInstr *DefInst = MRI.getVRegDef(Reg);
+          assert(DefInst);
+          MachineBasicBlock *RegBlock = DefInst->getParent();
+          Phi.addReg(Reg);
+          Phi.addMBB(RegBlock);
+          MBB.removeLiveIn(Reg);
+        }
+        RegisterAddressMap[PhiDstReg] = Address;
+        LiveAddressRegisterMap[Address] = PhiDstReg;
+      }
+      LI = MBB.livein_begin();
+    }
+
+    for (MachineBasicBlock::iterator I = MBB.begin(), Next = llvm::next(I);
+                               I != MBB.end(); I = Next) {
+      Next = llvm::next(I);
+      MachineInstr &MI = *I;
+
+      if (!TII->isRegisterLoad(MI)) {
+        if (MI.getOpcode() == AMDGPU::PHI) {
+          continue;
+        }
+        // Check for indirect register defs
+        for (unsigned OpIdx = 0, NumOperands = MI.getNumOperands();
+                                 OpIdx < NumOperands; ++OpIdx) {
+          MachineOperand &MO = MI.getOperand(OpIdx);
+          if (MO.isReg() && MO.isDef() &&
+              RegisterAddressMap.find(MO.getReg()) != RegisterAddressMap.end()) {
+            unsigned Reg = MO.getReg();
+            unsigned LiveAddress = RegisterAddressMap[Reg];
+            // Chain the live-ins
+            if (LiveAddressRegisterMap.find(LiveAddress) !=
+                                                     RegisterAddressMap.end()) {
+              MI.addOperand(MachineOperand::CreateReg(
+                                  LiveAddressRegisterMap[LiveAddress],
+                                  false, // isDef
+                                  true,  // isImp
+                                  true));  // isKill
+            }
+            LiveAddressRegisterMap[LiveAddress] = Reg;
+          }
+        }
+        continue;
+      }
+
+      const TargetRegisterClass *SuperIndirectRegClass =
+                                                TII->getSuperIndirectRegClass();
+      const TargetRegisterClass *IndirectLoadRegClass =
+                                             TII->getIndirectAddrLoadRegClass();
+      unsigned IndirectReg = MRI.createVirtualRegister(SuperIndirectRegClass);
+
+      unsigned RegIndex = MI.getOperand(2).getImm();
+      unsigned Channel = MI.getOperand(3).getImm();
+      unsigned Address = TII->calculateIndirectAddress(RegIndex, Channel);
+
+      if (MI.getOperand(1).getReg() == AMDGPU::INDIRECT_BASE_ADDR) {
+        // Direct register access
+        unsigned Reg = LiveAddressRegisterMap[Address];
+        unsigned AddrReg = IndirectLoadRegClass->getRegister(Address);
+
+        if (regHasExplicitDef(MRI, Reg)) {
+          // If the register we are reading from has an explicit def, then that
+          // means it was written via a direct register access (i.e. COPY
+          // or other instruction that doesn't use indirect addressing).  In
+          // this case we know where the value has been stored, so we can just
+          // issue a copy.
+          BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::COPY),
+                  MI.getOperand(0).getReg())
+                  .addReg(Reg);
+        } else {
+          // If the register we are reading has an implicit def, then that
+          // means it was written by an indirect register access (i.e. An
+          // instruction that uses indirect addressing. 
+          BuildMI(MBB, I, MBB.findDebugLoc(I), TII->get(AMDGPU::COPY),
+                   MI.getOperand(0).getReg())
+                   .addReg(AddrReg);
+        }
+      } else {
+        // Indirect register access
+
+        // Note on REQ_SEQUENCE instructons: You can't actually use the register
+        // it defines unless  you have an instruction that takes the defined
+        // register class as an operand.
+
+        MachineInstrBuilder Sequence = BuildMI(MBB, I, MBB.findDebugLoc(I),
+                                               TII->get(AMDGPU::REG_SEQUENCE),
+                                               IndirectReg);
+        for (int i = IndirectBegin; i <= IndirectEnd; ++i) {
+          unsigned Addr = TII->calculateIndirectAddress(i, Channel);
+          if (LiveAddressRegisterMap.find(Addr) == LiveAddressRegisterMap.end()) {
+            continue;
+          }
+          unsigned Reg = LiveAddressRegisterMap[Addr];
+
+          // We only need to use REG_SEQUENCE for explicit defs, since the
+          // register coalescer won't do anything with the implicit defs.
+          MachineInstr *DefInstr = MRI.getVRegDef(Reg);
+          if (!DefInstr->getOperand(0).isReg() ||
+              DefInstr->getOperand(0).getReg() != Reg) {
+            continue;
+          }
+
+          // Insert a REQ_SEQUENCE instruction to force the register allocator
+          // to allocate the virtual register to the correct physical register.
+          Sequence.addReg(LiveAddressRegisterMap[Addr]);
+          Sequence.addImm(TII->getRegisterInfo().getIndirectSubReg(Addr));
+        }
+        MachineInstrBuilder Mov = TII->buildIndirectRead(BB, I,
+                                           MI.getOperand(0).getReg(), // Value
+                                           Address,
+                                           MI.getOperand(1).getReg()); // Offset
+
+
+
+        Mov.addReg(IndirectReg, RegState::Implicit | RegState::Kill);
+
+      }
+      MI.eraseFromParent();
+    }
+  }
+  return false;
+}
+
+bool AMDGPUIndirectAddressingPass::regHasExplicitDef(MachineRegisterInfo &MRI,
+                                                  unsigned Reg) const {
+  MachineInstr *DefInstr = MRI.getVRegDef(Reg);
+  return DefInstr && DefInstr->getOperand(0).isReg() &&
+         DefInstr->getOperand(0).getReg() == Reg;
+}
diff --git a/lib/Target/R600/AMDGPUInstrInfo.cpp b/lib/Target/R600/AMDGPUInstrInfo.cpp
index e42a46d839..640707d7ca 100644
--- a/lib/Target/R600/AMDGPUInstrInfo.cpp
+++ b/lib/Target/R600/AMDGPUInstrInfo.cpp
@@ -234,7 +234,16 @@ AMDGPUInstrInfo::isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const {
   // TODO: Implement this function
   return true;
 }
- 
+
+bool AMDGPUInstrInfo::isRegisterStore(const MachineInstr &MI) const {
+  return get(MI.getOpcode()).TSFlags & AMDGPU_FLAG_REGISTER_STORE;
+}
+
+bool AMDGPUInstrInfo::isRegisterLoad(const MachineInstr &MI) const {
+  return get(MI.getOpcode()).TSFlags & AMDGPU_FLAG_REGISTER_LOAD;
+}
+
+
 void AMDGPUInstrInfo::convertToISA(MachineInstr & MI, MachineFunction &MF,
     DebugLoc DL) const {
   MachineRegisterInfo &MRI = MF.getRegInfo();
diff --git a/lib/Target/R600/AMDGPUInstrInfo.h b/lib/Target/R600/AMDGPUInstrInfo.h
index cb97af9831..3909e4e105 100644
--- a/lib/Target/R600/AMDGPUInstrInfo.h
+++ b/lib/Target/R600/AMDGPUInstrInfo.h
@@ -40,9 +40,10 @@ class MachineInstrBuilder;
 class AMDGPUInstrInfo : public AMDGPUGenInstrInfo {
 private:
   const AMDGPURegisterInfo RI;
-  TargetMachine &TM;
   bool getNextBranchInstr(MachineBasicBlock::iterator &iter,
                           MachineBasicBlock &MBB) const;
+protected:
+  TargetMachine &TM;
 public:
   explicit AMDGPUInstrInfo(TargetMachine &tm);
 
@@ -130,12 +131,66 @@ public:
   bool isAExtLoadInst(llvm::MachineInstr *MI) const;
   bool isStoreInst(llvm::MachineInstr *MI) const;
   bool isTruncStoreInst(llvm::MachineInstr *MI) const;
+  bool isRegisterStore(const MachineInstr &MI) const;
+  bool isRegisterLoad(const MachineInstr &MI) const;
+
+//===---------------------------------------------------------------------===//
+// Pure virtual funtions to be implemented by sub-classes.
+//===---------------------------------------------------------------------===//
 
   virtual MachineInstr* getMovImmInstr(MachineFunction *MF, unsigned DstReg,
                                        int64_t Imm) const = 0;
   virtual unsigned getIEQOpcode() const = 0;
   virtual bool isMov(unsigned opcode) const = 0;
 
+  /// \returns the smallest register index that will be accessed by an indirect
+  /// read or write or -1 if indirect addressing is not used by this program.
+  virtual int getIndirectIndexBegin(const MachineFunction &MF) const = 0;
+
+  /// \returns the largest register index that will be accessed by an indirect
+  /// read or write or -1 if indirect addressing is not used by this program.
+  virtual int getIndirectIndexEnd(const MachineFunction &MF) const = 0;
+
+  /// \brief Calculate the "Indirect Address" for the given \p RegIndex and
+  ///        \p Channel
+  ///
+  /// We model indirect addressing using a virtual address space that can be
+  /// accesed with loads and stores.  The "Indirect Address" is the memory
+  /// address in this virtual address space that maps to the given \p RegIndex
+  /// and \p Channel.
+  virtual unsigned calculateIndirectAddress(unsigned RegIndex,
+                                            unsigned Channel) const = 0;
+
+  /// \returns The register class to be used for storing values to an
+  /// "Indirect Address" .
+  virtual const TargetRegisterClass *getIndirectAddrStoreRegClass(
+                                                  unsigned SourceReg) const = 0;
+
+  /// \returns The register class to be used for loading values from
+  /// an "Indirect Address" .
+  virtual const TargetRegisterClass *getIndirectAddrLoadRegClass() const = 0;
+
+  /// \brief Build instruction(s) for an indirect register write.
+  ///
+  /// \returns The instruction that performs the indirect register write
+  virtual MachineInstrBuilder buildIndirectWrite(MachineBasicBlock *MBB,
+                                    MachineBasicBlock::iterator I,
+                                    unsigned ValueReg, unsigned Address,
+                                    unsigned OffsetReg) const = 0;
+
+  /// \brief Build instruction(s) for an indirect register read.
+  ///
+  /// \returns The instruction that performs the indirect register read
+  virtual MachineInstrBuilder buildIndirectRead(MachineBasicBlock *MBB,
+                                    MachineBasicBlock::iterator I,
+                                    unsigned ValueReg, unsigned Address,
+                                    unsigned OffsetReg) const = 0;
+
+  /// \returns the register class whose sub registers are the set of all
+  /// possible registers that can be used for indirect addressing.
+  virtual const TargetRegisterClass *getSuperIndirectRegClass() const = 0;
+
+
   /// \brief Convert the AMDIL MachineInstr to a supported ISA
   /// MachineInstr
   virtual void convertToISA(MachineInstr & MI, MachineFunction &MF,
@@ -145,4 +200,7 @@ public:
 
 } // End llvm namespace
 
+#define AMDGPU_FLAG_REGISTER_LOAD  (UINT64_C(1) << 63)
+#define AMDGPU_FLAG_REGISTER_STORE (UINT64_C(1) << 62)
+
 #endif // AMDGPUINSTRINFO_H
diff --git a/lib/Target/R600/AMDGPUInstrInfo.td b/lib/Target/R600/AMDGPUInstrInfo.td
index 96368e8541..b66ae879dc 100644
--- a/lib/Target/R600/AMDGPUInstrInfo.td
+++ b/lib/Target/R600/AMDGPUInstrInfo.td
@@ -72,3 +72,11 @@ def AMDGPUumin : SDNode<"AMDGPUISD::UMIN", SDTIntBinOp,
 def AMDGPUurecip : SDNode<"AMDGPUISD::URECIP", SDTIntUnaryOp>;
 
 def fpow : SDNode<"ISD::FPOW", SDTFPBinOp>;
+
+def AMDGPUregister_load : SDNode<"AMDGPUISD::REGISTER_LOAD",
+                          SDTypeProfile<1, 2, [SDTCisPtrTy<1>, SDTCisInt<2>]>,
+                          [SDNPHasChain, SDNPMayLoad]>;
+
+def AMDGPUregister_store : SDNode<"AMDGPUISD::REGISTER_STORE",
+                           SDTypeProfile<0, 3, [SDTCisPtrTy<1>, SDTCisInt<2>]>,
+                           [SDNPHasChain, SDNPMayStore]>;
diff --git a/lib/Target/R600/AMDGPUInstructions.td b/lib/Target/R600/AMDGPUInstructions.td
index e634d20b61..3dee004270 100644
--- a/lib/Target/R600/AMDGPUInstructions.td
+++ b/lib/Target/R600/AMDGPUInstructions.td
@@ -13,8 +13,8 @@
 //===----------------------------------------------------------------------===//
 
 class AMDGPUInst <dag outs, dag ins, string asm, list<dag> pattern> : Instruction {
-  field bits<16> AMDILOp = 0;
-  field bits<3> Gen = 0;
+  field bit isRegisterLoad = 0;
+  field bit isRegisterStore = 0;
 
   let Namespace = "AMDGPU";
   let OutOperandList = outs;
@@ -22,8 +22,9 @@ class AMDGPUInst <dag outs, dag ins, string asm, list<dag> pattern> : Instructio
   let AsmString = asm;
   let Pattern = pattern;
   let Itinerary = NullALU;
-  let TSFlags{42-40} = Gen;
-  let TSFlags{63-48} = AMDILOp;
+
+  let TSFlags{63} = isRegisterLoad;
+  let TSFlags{62} = isRegisterStore;
 }
 
 class AMDGPUShaderInst <dag outs, dag ins, string asm, list<dag> pattern>
@@ -101,7 +102,9 @@ def FP_ONE : PatLeaf <
   [{return N->isExactlyValue(1.0);}]
 >;
 
-let isCodeGenOnly = 1, isPseudo = 1, usesCustomInserter = 1  in {
+let isCodeGenOnly = 1, isPseudo = 1 in {
+
+let usesCustomInserter = 1  in {