diff --git a/lib/Target/R600/AMDGPU.h b/lib/Target/R600/AMDGPU.h
index e099a9fc319..3cd792ad4aa 100644
--- a/lib/Target/R600/AMDGPU.h
+++ b/lib/Target/R600/AMDGPU.h
@@ -23,6 +23,7 @@ class AMDGPUTargetMachine;
 // R600 Passes
 FunctionPass* createR600KernelParametersPass(const DataLayout *TD);
 FunctionPass *createR600ExpandSpecialInstrsPass(TargetMachine &tm);
+FunctionPass *createR600EmitClauseMarkers(TargetMachine &tm);
 
 // SI Passes
 FunctionPass *createSIAnnotateControlFlowPass();
diff --git a/lib/Target/R600/AMDGPUTargetMachine.cpp b/lib/Target/R600/AMDGPUTargetMachine.cpp
index 01857475440..45b1be081c2 100644
--- a/lib/Target/R600/AMDGPUTargetMachine.cpp
+++ b/lib/Target/R600/AMDGPUTargetMachine.cpp
@@ -151,6 +151,7 @@ bool AMDGPUPassConfig::addPreEmitPass() {
   if (ST.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX) {
     addPass(createAMDGPUCFGPreparationPass(*TM));
     addPass(createAMDGPUCFGStructurizerPass(*TM));
+    addPass(createR600EmitClauseMarkers(*TM));
     addPass(createR600ExpandSpecialInstrsPass(*TM));
     addPass(&FinalizeMachineBundlesID);
   } else {
diff --git a/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp b/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
index a945fe9ef81..1bf87fc0963 100644
--- a/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
+++ b/lib/Target/R600/MCTargetDesc/R600MCCodeEmitter.cpp
@@ -101,7 +101,8 @@ enum InstrTypes {
   INSTR_FC,
   INSTR_NATIVE,
   INSTR_VTX,
-  INSTR_EXPORT
+  INSTR_EXPORT,
+  INSTR_CFALU
 };
 
 enum FCInstr {
@@ -251,6 +252,13 @@ void R600MCCodeEmitter::EncodeInstruction(const MCInst &MI, raw_ostream &OS,
       Emit(Inst, OS);
       break;
     }
+    case AMDGPU::CF_ALU:
+    case AMDGPU::CF_ALU_PUSH_BEFORE: {
+      uint64_t Inst = getBinaryCodeForInstr(MI, Fixups);
+      EmitByte(INSTR_CFALU, OS);
+      Emit(Inst, OS);
+      break;
+    }
 
     default:
       EmitALUInstr(MI, Fixups, OS);
diff --git a/lib/Target/R600/R600EmitClauseMarkers.cpp b/lib/Target/R600/R600EmitClauseMarkers.cpp
new file mode 100644
index 00000000000..7c7469a04b2
--- /dev/null
+++ b/lib/Target/R600/R600EmitClauseMarkers.cpp
@@ -0,0 +1,253 @@
+//===-- R600EmitClauseMarkers.cpp - Emit CF_ALU ---------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// Add CF_ALU. R600 Alu instructions are grouped in clause which can hold
+/// 128 Alu instructions ; these instructions can access up to 4 prefetched
+/// 4 lines of 16 registers from constant buffers. Such ALU clauses are
+/// initiated by CF_ALU instructions.
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "R600Defines.h"
+#include "R600InstrInfo.h"
+#include "R600MachineFunctionInfo.h"
+#include "R600RegisterInfo.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+
+namespace llvm {
+
+class R600EmitClauseMarkersPass : public MachineFunctionPass {
+
+private:
+  static char ID;
+  const R600InstrInfo *TII;
+
+  unsigned OccupiedDwords(MachineInstr *MI) const {
+    switch (MI->getOpcode()) {
+    case AMDGPU::INTERP_PAIR_XY:
+    case AMDGPU::INTERP_PAIR_ZW:
+    case AMDGPU::INTERP_VEC_LOAD:
+    case AMDGPU::DOT4_eg_pseudo:
+    case AMDGPU::DOT4_r600_pseudo:
+      return 4;
+    case AMDGPU::KILL:
+      return 0;
+    default:
+      break;
+    }
+
+    if(TII->isVector(*MI) ||
+        TII->isCubeOp(MI->getOpcode()) ||
+        TII->isReductionOp(MI->getOpcode()))
+      return 4;
+
+    unsigned NumLiteral = 0;
+    for (MachineInstr::mop_iterator It = MI->operands_begin(),
+        E = MI->operands_end(); It != E; ++It) {
+      MachineOperand &MO = *It;
+      if (MO.isReg() && MO.getReg() == AMDGPU::ALU_LITERAL_X)
+        ++NumLiteral;
+    }
+    return 1 + NumLiteral;
+  }
+
+  bool isALU(const MachineInstr *MI) const {
+    if (MI->getOpcode() == AMDGPU::KILLGT)
+      return false;
+    if (TII->isALUInstr(MI->getOpcode()))
+      return true;
+    if (TII->isVector(*MI) || TII->isCubeOp(MI->getOpcode()))
+      return true;
+    switch (MI->getOpcode()) {
+    case AMDGPU::PRED_X:
+    case AMDGPU::INTERP_PAIR_XY:
+    case AMDGPU::INTERP_PAIR_ZW:
+    case AMDGPU::INTERP_VEC_LOAD:
+    case AMDGPU::COPY:
+    case AMDGPU::DOT4_eg_pseudo:
+    case AMDGPU::DOT4_r600_pseudo:
+      return true;
+    default:
+      return false;
+    }
+  }
+
+  bool IsTrivialInst(MachineInstr *MI) const {
+    switch (MI->getOpcode()) {
+    case AMDGPU::KILL:
+    case AMDGPU::RETURN:
+      return true;
+    default:
+      return false;
+    }
+  }
+
+  // Register Idx, then Const value
+  std::vector<std::pair<unsigned, unsigned> > ExtractConstRead(MachineInstr *MI)
+      const {
+    const R600Operands::Ops OpTable[3][2] = {
+      {R600Operands::SRC0, R600Operands::SRC0_SEL},
+      {R600Operands::SRC1, R600Operands::SRC1_SEL},
+      {R600Operands::SRC2, R600Operands::SRC2_SEL},
+    };
+    std::vector<std::pair<unsigned, unsigned> > Result;
+
+    if (!TII->isALUInstr(MI->getOpcode()))
+      return Result;
+    for (unsigned j = 0; j < 3; j++) {
+      int SrcIdx = TII->getOperandIdx(MI->getOpcode(), OpTable[j][0]);
+      if (SrcIdx < 0)
+        break;
+      if (MI->getOperand(SrcIdx).getReg() == AMDGPU::ALU_CONST) {
+        unsigned Const = MI->getOperand(
+            TII->getOperandIdx(MI->getOpcode(), OpTable[j][1])).getImm();
+        Result.push_back(std::pair<unsigned, unsigned>(SrcIdx, Const));
+      }
+    }
+    return Result;
+  }
+
+  std::pair<unsigned, unsigned> getAccessedBankLine(unsigned Sel) const {
+    // Sel is (512 + (kc_bank << 12) + ConstIndex) << 2
+    // (See also R600ISelLowering.cpp)
+    // ConstIndex value is in [0, 4095];
+    return std::pair<unsigned, unsigned>(
+        ((Sel >> 2) - 512) >> 12, // KC_BANK
+        // Line Number of ConstIndex
+        // A line contains 16 constant registers however KCX bank can lock
+        // two line at the same time ; thus we want to get an even line number.
+        // Line number can be retrieved with (>>4), using (>>5) <<1 generates
+        // an even number.
+        ((((Sel >> 2) - 512) & 4095) >> 5) << 1);
+  }
+
+  bool SubstituteKCacheBank(MachineInstr *MI,
+      std::vector<std::pair<unsigned, unsigned> > &CachedConsts) const {
+    std::vector<std::pair<unsigned, unsigned> > UsedKCache;
+    std::vector<std::pair<unsigned, unsigned> > Consts = ExtractConstRead(MI);
+    assert(TII->isALUInstr(MI->getOpcode()) && "Can't assign Const");
+    for (unsigned i = 0, n = Consts.size(); i < n; ++i) {
+      unsigned Sel = Consts[i].second;
+      unsigned Chan = Sel & 3, Index = ((Sel >> 2) - 512) & 31;
+      unsigned KCacheIndex = Index * 4 + Chan;
+      const std::pair<unsigned, unsigned> &BankLine = getAccessedBankLine(Sel);
+      if (CachedConsts.empty()) {
+        CachedConsts.push_back(BankLine);
+        UsedKCache.push_back(std::pair<unsigned, unsigned>(0, KCacheIndex));
+        continue;
+      }
+      if (CachedConsts[0] == BankLine) {
+        UsedKCache.push_back(std::pair<unsigned, unsigned>(0, KCacheIndex));
+        continue;
+      }
+      if (CachedConsts.size() == 1) {
+        CachedConsts.push_back(BankLine);
+        UsedKCache.push_back(std::pair<unsigned, unsigned>(1, KCacheIndex));
+        continue;
+      }
+      if (CachedConsts[1] == BankLine) {
+        UsedKCache.push_back(std::pair<unsigned, unsigned>(1, KCacheIndex));
+        continue;
+      }
+      return false;
+    }
+
+    for (unsigned i = 0, n = Consts.size(); i < n; ++i) {
+      switch(UsedKCache[i].first) {
+      case 0:
+        MI->getOperand(Consts[i].first).setReg(
+            AMDGPU::R600_KC0RegClass.getRegister(UsedKCache[i].second));
+        break;
+      case 1:
+        MI->getOperand(Consts[i].first).setReg(
+            AMDGPU::R600_KC1RegClass.getRegister(UsedKCache[i].second));
+        break;
+      default:
+        llvm_unreachable("Wrong Cache Line");
+      }
+    }
+    return true;
+  }
+
+  MachineBasicBlock::iterator
+  MakeALUClause(MachineBasicBlock &MBB, MachineBasicBlock::iterator I) const {
+    MachineBasicBlock::iterator ClauseHead = I;
+    std::vector<std::pair<unsigned, unsigned> > KCacheBanks;
+    bool PushBeforeModifier = false;
+    unsigned AluInstCount = 0;
+    for (MachineBasicBlock::iterator E = MBB.end(); I != E; ++I) {
+      if (IsTrivialInst(I))
+        continue;
+      if (!isALU(I))
+        break;
+      if (I->getOpcode() == AMDGPU::PRED_X) {
+        if (TII->getFlagOp(I).getImm() & MO_FLAG_PUSH)
+          PushBeforeModifier = true;
+        AluInstCount ++;
+        continue;
+      }
+      if (TII->isALUInstr(I->getOpcode()) &&
+          !SubstituteKCacheBank(I, KCacheBanks))
+        break;
+      AluInstCount += OccupiedDwords(I);
+      if (AluInstCount > 124)
+        break;
+    }
+    unsigned Opcode = PushBeforeModifier ?
+        AMDGPU::CF_ALU_PUSH_BEFORE : AMDGPU::CF_ALU;
+    BuildMI(MBB, ClauseHead, MBB.findDebugLoc(ClauseHead), TII->get(Opcode))
+        .addImm(0) // ADDR
+        .addImm(KCacheBanks.empty()?0:KCacheBanks[0].first) // KB0
+        .addImm((KCacheBanks.size() < 2)?0:KCacheBanks[1].first) // KB1
+        .addImm(KCacheBanks.empty()?0:2) // KM0
+        .addImm((KCacheBanks.size() < 2)?0:2) // KM1
+        .addImm(KCacheBanks.empty()?0:KCacheBanks[0].second) // KLINE0
+        .addImm((KCacheBanks.size() < 2)?0:KCacheBanks[1].second) // KLINE1
+        .addImm(AluInstCount); // COUNT
+    return I;
+  }
+
+public:
+  R600EmitClauseMarkersPass(TargetMachine &tm) : MachineFunctionPass(ID),
+    TII (static_cast<const R600InstrInfo *>(tm.getInstrInfo())) { }
+
+  virtual bool runOnMachineFunction(MachineFunction &MF) {
+    for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
+                                                    BB != BB_E; ++BB) {
+      MachineBasicBlock &MBB = *BB;
+      MachineBasicBlock::iterator I = MBB.begin();
+      if (I->getOpcode() == AMDGPU::CF_ALU)
+        continue; // BB was already parsed
+      for (MachineBasicBlock::iterator E = MBB.end(); I != E;) {
+        if (isALU(I))
+          I = MakeALUClause(MBB, I);
+        else
+          ++I;
+      }
+    }
+    return false;
+  }
+
+  const char *getPassName() const {
+    return "R600 Emit Clause Markers Pass";
+  }
+};
+
+char R600EmitClauseMarkersPass::ID = 0;
+
+}
+
+
+llvm::FunctionPass *llvm::createR600EmitClauseMarkers(TargetMachine &TM) {
+  return new R600EmitClauseMarkersPass(TM);
+}
+
diff --git a/lib/Target/R600/R600Instructions.td b/lib/Target/R600/R600Instructions.td
index 7550ecec6fc..7cfde78fe5d 100644
--- a/lib/Target/R600/R600Instructions.td
+++ b/lib/Target/R600/R600Instructions.td
@@ -351,9 +351,9 @@ class R600_1OP <bits<11> inst, string opName, list<dag> pattern,
               (ins WRITE:$write, OMOD:$omod, REL:$dst_rel, CLAMP:$clamp,
                    R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
                    LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal),
-              !strconcat(opName,
+              !strconcat("  ", opName,
                    "$clamp $dst$write$dst_rel$omod, "
-                   "$src0_neg$src0_abs$src0$src0_sel$src0_abs$src0_rel, "
+                   "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
                    "$literal $pred_sel$last"),
               pattern,
               itin>,
@@ -392,10 +392,10 @@ class R600_2OP <bits<11> inst, string opName, list<dag> pattern,
                R600_Reg32:$src0, NEG:$src0_neg, REL:$src0_rel, ABS:$src0_abs, SEL:$src0_sel,
                R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, ABS:$src1_abs, SEL:$src1_sel,
                LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal),
-          !strconcat(opName,
+          !strconcat("  ", opName,
                 "$clamp $update_exec_mask$update_pred$dst$write$dst_rel$omod, "
-                "$src0_neg$src0_abs$src0$src0_sel$src0_abs$src0_rel, "
-                "$src1_neg$src1_abs$src1$src1_sel$src1_abs$src1_rel, "
+                "$src0_neg$src0_abs$src0$src0_abs$src0_rel, "
+                "$src1_neg$src1_abs$src1$src1_abs$src1_rel, "
                 "$literal $pred_sel$last"),
           pattern,
           itin>,
@@ -430,10 +430,10 @@ class R600_3OP <bits<5> inst, string opName, list<dag> pattern,
                R600_Reg32:$src1, NEG:$src1_neg, REL:$src1_rel, SEL:$src1_sel,
                R600_Reg32:$src2, NEG:$src2_neg, REL:$src2_rel, SEL:$src2_sel,
                LAST:$last, R600_Pred:$pred_sel, LITERAL:$literal),
-          !strconcat(opName, "$clamp $dst$dst_rel, "
-                             "$src0_neg$src0$src0_sel$src0_rel, "
-                             "$src1_neg$src1$src1_sel$src1_rel, "
-                             "$src2_neg$src2$src2_sel$src2_rel, "
+          !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>,
@@ -765,6 +765,67 @@ class ExportBufInst : InstR600ISA<(
   let Inst{63-32} = Word1;
 }
 
+//===----------------------------------------------------------------------===//
+// Control Flow Instructions
+//===----------------------------------------------------------------------===//
+
+class CF_ALU_WORD0 {
+  field bits<32> Word0;
+
+  bits<22> ADDR;
+  bits<4> KCACHE_BANK0;
+  bits<4> KCACHE_BANK1;
+  bits<2> KCACHE_MODE0;
+
+  let Word0{21-0} = ADDR;
+  let Word0{25-22} = KCACHE_BANK0;
+  let Word0{29-26} = KCACHE_BANK1;
+  let Word0{31-30} = KCACHE_MODE0;
+}
+
+class CF_ALU_WORD1 {
+  field bits<32> Word1;
+
+  bits<2> KCACHE_MODE1;
+  bits<8> KCACHE_ADDR0;
+  bits<8> KCACHE_ADDR1;
+  bits<7> COUNT;
+  bits<1> ALT_CONST;
+  bits<4> CF_INST;
+  bits<1> WHOLE_QUAD_MODE;
+  bits<1> BARRIER;
+
+  let Word1{1-0} = KCACHE_MODE1;
+  let Word1{9-2} = KCACHE_ADDR0;
+  let Word1{17-10} = KCACHE_ADDR1;
+  let Word1{24-18} = COUNT;
+  let Word1{25} = ALT_CONST;
+  let Word1{29-26} = CF_INST;
+  let Word1{30} = WHOLE_QUAD_MODE;
+  let Word1{31} = BARRIER;
+}
+
+class ALU_CLAUSE<bits<4> inst, string OpName> : AMDGPUInst <(outs),
+(ins i32imm:$ADDR, i32imm:$KCACHE_BANK0, i32imm:$KCACHE_BANK1, i32imm:$KCACHE_MODE0, i32imm:$KCACHE_MODE1,
+i32imm:$KCACHE_ADDR0, i32imm:$KCACHE_ADDR1, i32imm:$COUNT),
+!strconcat(OpName, " $COUNT, @$ADDR, "
+"KC0[CB$KCACHE_BANK0:$KCACHE_ADDR0-$KCACHE_ADDR0+32]"
+", KC1[CB$KCACHE_BANK1:$KCACHE_ADDR1-$KCACHE_ADDR1+32]"),
+[] >, CF_ALU_WORD0, CF_ALU_WORD1 {
+  field bits<64> Inst;
+
+  let CF_INST = inst;
+  let ALT_CONST = 0;
+  let WHOLE_QUAD_MODE = 0;
+  let BARRIER = 1;
+
+  let Inst{31-0} = Word0;
+  let Inst{63-32} = Word1;
+}
+
+def CF_ALU : ALU_CLAUSE<8, "ALU">;
+def CF_ALU_PUSH_BEFORE : ALU_CLAUSE<9, "ALU_PUSH_BEFORE">;
+
 let Predicates = [isR600toCayman] in {
 
 //===----------------------------------------------------------------------===//
diff --git a/lib/Target/R600/R600RegisterInfo.td b/lib/Target/R600/R600RegisterInfo.td
index ce5994ca368..03f49761ea4 100644
--- a/lib/Target/R600/R600RegisterInfo.td
+++ b/lib/Target/R600/R600RegisterInfo.td
@@ -43,6 +43,37 @@ foreach Index = 0-127 in {
                                    Index>;
 }
 
+// KCACHE_BANK0
+foreach Index = 159-128 in {
+  foreach Chan = [ "X", "Y", "Z", "W" ] in {
+    // 32-bit Temporary Registers
+    def KC0_#Index#_#Chan : R600RegWithChan <"KC0["#Index#"-128]."#Chan, Index, Chan>;
+  }
+  // 128-bit Temporary Registers
+  def KC0_#Index#_XYZW : R600Reg_128 <"KC0["#Index#"-128].XYZW",
+                                 [!cast<Register>("KC0_"#Index#"_X"),
+                                  !cast<Register>("KC0_"#Index#"_Y"),
+                                  !cast<Register>("KC0_"#Index#"_Z"),
+                                  !cast<Register>("KC0_"#Index#"_W")],
+                                 Index>;
+}
+
+// KCACHE_BANK1
+foreach Index = 191-160 in {
+  foreach Chan = [ "X", "Y", "Z", "W" ] in {
+    // 32-bit Temporary Registers
+    def KC1_#Index#_#Chan : R600RegWithChan <"KC1["#Index#"-160]."#Chan, Index, Chan>;
+  }
+  // 128-bit Temporary Registers
+  def KC1_#Index#_XYZW : R600Reg_128 <"KC1["#Index#"-160].XYZW",
+                                 [!cast<Register>("KC1_"#Index#"_X"),
+                                  !cast<Register>("KC1_"#Index#"_Y"),
+                                  !cast<Register>("KC1_"#Index#"_Z"),
+                                  !cast<Register>("KC1_"#Index#"_W")],
+                                 Index>;
+}
+
+
 // Array Base Register holding input in FS
 foreach Index = 448-480 in {
   def ArrayBase#Index :  R600Reg<"ARRAY_BASE", Index>;
@@ -80,6 +111,38 @@ def R600_Addr : RegisterClass <"AMDGPU", [i32], 127, (add (sequence "Addr%u_X",
 
 } // End isAllocatable = 0
 
+def R600_KC0_X : RegisterClass <"AMDGPU", [f32, i32], 32,
+                              (add (sequence "KC0_%u_X", 128, 159))>;
+
+def R600_KC0_Y : RegisterClass <"AMDGPU", [f32, i32], 32,
+                              (add (sequence "KC0_%u_Y", 128, 159))>;
+
+def R600_KC0_Z : RegisterClass <"AMDGPU", [f32, i32], 32,
+                              (add (sequence "KC0_%u_Z", 128, 159))>;
+
+def R600_KC0_W : RegisterClass <"AMDGPU", [f32, i32], 32,
+                              (add (sequence "KC0_%u_W", 128, 159))>;
+
+def R600_KC0 : RegisterClass <"AMDGPU", [f32, i32], 32,
+                                   (interleave R600_KC0_X, R600_KC0_Y,
+                                               R600_KC0_Z, R600_KC0_W)>;
+
+def R600_KC1_X : RegisterClass <"AMDGPU", [f32, i32], 32,
+                              (add (sequence "KC1_%u_X", 160, 191))>;
+
+def R600_KC1_Y : RegisterClass <"AMDGPU", [f32, i32], 32,
+                              (add (sequence "KC1_%u_Y", 160, 191))>;
+
+def R600_KC1_Z : RegisterClass <"AMDGPU", [f32, i32], 32,
+                              (add (sequence "KC1_%u_Z", 160, 191))>;
+
+def R600_KC1_W : RegisterClass <"AMDGPU", [f32, i32], 32,
+                              (add (sequence "KC1_%u_W", 160, 191))>;
+
+def R600_KC1 : RegisterClass <"AMDGPU", [f32, i32], 32,
+                                   (interleave R600_KC1_X, R600_KC1_Y,
+                                               R600_KC1_Z, R600_KC1_W)>;
+
 def R600_TReg32_X : RegisterClass <"AMDGPU", [f32, i32], 32,
                                    (add (sequence "T%u_X", 0, 127), AR_X)>;
 
diff --git a/test/CodeGen/R600/kcache-fold.ll b/test/CodeGen/R600/kcache-fold.ll
index e8e2bf5f0f6..3d70e4bd54a 100644
--- a/test/CodeGen/R600/kcache-fold.ll
+++ b/test/CodeGen/R600/kcache-fold.ll
@@ -1,7 +1,7 @@
 ;RUN: llc < %s -march=r600 -mcpu=redwood | FileCheck %s
 
 ; CHECK: @main1
-; CHECK: MOV T{{[0-9]+\.[XYZW], CBuf0\[[0-9]+\]\.[XYZW]}}
+; CHECK: MOV T{{[0-9]+\.[XYZW], KC0}}
 define void @main1() {
 main_body:
   %0 = load <4 x float> addrspace(8)* null