[Hexagon] Adding functionality for duplexing. Duplexing is a way to compress commonly used pairs of instructions in order to reduce code size. The test case duplex.ll normally would be 8 bytes, assign register to 0 and jump to link register. After duplexing this is only 4 bytes. This also tests the HexagonMCShuffler code path which is used to make sure duplexed instructions still follow slot requirements.

llvm-svn: 239095

9 files changed, 1383 insertions, 4 deletions

diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/CMakeLists.txt b/llvm/lib/Target/Hexagon/MCTargetDesc/CMakeLists.txt
index 90b34d304b7..d0fae8a5192 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/CMakeLists.txt
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/CMakeLists.txt
@@ -4,6 +4,7 @@ add_llvm_library(LLVMHexagonDesc
   HexagonInstPrinter.cpp
   HexagonMCAsmInfo.cpp
   HexagonMCCodeEmitter.cpp
+  HexagonMCDuplexInfo.cpp
   HexagonMCInstrInfo.cpp
   HexagonMCShuffler.cpp
   HexagonMCTargetDesc.cpp
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonAsmBackend.cpp b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonAsmBackend.cpp
index b4db977e361..76894840153 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonAsmBackend.cpp
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonAsmBackend.cpp
@@ -16,6 +16,7 @@
 #include "llvm/MC/MCAssembler.h"
 #include "llvm/MC/MCELFObjectWriter.h"
 #include "llvm/MC/MCFixupKindInfo.h"
+#include "llvm/MC/MCInstrInfo.h"
 #include "llvm/Support/TargetRegistry.h"
 
 using namespace llvm;
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonBaseInfo.h b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonBaseInfo.h
index 342a61190e6..4f84e0d79e6 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonBaseInfo.h
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonBaseInfo.h
@@ -43,6 +43,7 @@ namespace HexagonII {
     TypeXTYPE   = 8,
     TypeMEMOP   = 9,
     TypeNV      = 10,
+    TypeDUPLEX  = 11,
     TypePREFIX  = 30, // Such as extenders.
     TypeENDLOOP = 31  // Such as end of a HW loop.
   };
@@ -190,6 +191,17 @@ namespace HexagonII {
     MO_GPREL
   };
 
+  // Hexagon Sub-instruction classes.
+  enum SubInstructionGroup {
+    HSIG_None = 0,
+    HSIG_L1,
+    HSIG_L2,
+    HSIG_S1,
+    HSIG_S2,
+    HSIG_A,
+    HSIG_Compound
+  };
+
   enum InstParseBits {
     INST_PARSE_MASK       = 0x0000c000,
     INST_PARSE_PACKET_END = 0x0000c000,
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonInstPrinter.cpp b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonInstPrinter.cpp
index 73f20f52439..3fe20395c40 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonInstPrinter.cpp
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonInstPrinter.cpp
@@ -129,7 +129,13 @@ void HexagonInstPrinter::printInst(MCInst const *MI, raw_ostream &OS,
   HasExtender = false;
   for (auto const &I : HexagonMCInstrInfo::bundleInstructions(*MI)) {
     MCInst const &MCI = *I.getInst();
-    printInstruction(&MCI, OS);
+    if (HexagonMCInstrInfo::isDuplex(MII, MCI)) {
+      printInstruction(MCI.getOperand(1).getInst(), OS);
+      OS << '\v';
+      HasExtender = false;
+      printInstruction(MCI.getOperand(0).getInst(), OS);
+    } else
+      printInstruction(&MCI, OS);
     setExtender(MCI);
     OS << "\n";
   }
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCCodeEmitter.cpp b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCCodeEmitter.cpp
index 6796a22f725..1eee852996f 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCCodeEmitter.cpp
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCCodeEmitter.cpp
@@ -40,18 +40,25 @@ HexagonMCCodeEmitter::HexagonMCCodeEmitter(MCInstrInfo const &aMII,
 uint32_t HexagonMCCodeEmitter::parseBits(size_t Instruction, size_t Last,
                                          MCInst const &MCB,
                                          MCInst const &MCI) const {
+  bool Duplex = HexagonMCInstrInfo::isDuplex(MCII, MCI);
   if (Instruction == 0) {
     if (HexagonMCInstrInfo::isInnerLoop(MCB)) {
+      assert(!Duplex);
       assert(Instruction != Last);
       return HexagonII::INST_PARSE_LOOP_END;
     }
   }
   if (Instruction == 1) {
     if (HexagonMCInstrInfo::isOuterLoop(MCB)) {
+      assert(!Duplex);
       assert(Instruction != Last);
       return HexagonII::INST_PARSE_LOOP_END;
     }
   }
+  if (Duplex) {
+    assert(Instruction == Last);
+    return HexagonII::INST_PARSE_DUPLEX;
+  }
   if(Instruction == Last)
     return HexagonII::INST_PARSE_PACKET_END;
   return HexagonII::INST_PARSE_NOT_END;
@@ -149,6 +156,81 @@ void HexagonMCCodeEmitter::EncodeSingleInstruction(
     llvm_unreachable("Unimplemented Instruction");
   }
   Binary |= Parse;
+
+  // if we need to emit a duplexed instruction
+  if (HMB.getOpcode() >= Hexagon::DuplexIClass0 &&
+      HMB.getOpcode() <= Hexagon::DuplexIClassF) {
+    assert(Parse == HexagonII::INST_PARSE_DUPLEX &&
+           "Emitting duplex without duplex parse bits");
+    unsigned dupIClass;
+    switch (HMB.getOpcode()) {
+    case Hexagon::DuplexIClass0:
+      dupIClass = 0;
+      break;
+    case Hexagon::DuplexIClass1:
+      dupIClass = 1;
+      break;
+    case Hexagon::DuplexIClass2:
+      dupIClass = 2;
+      break;
+    case Hexagon::DuplexIClass3:
+      dupIClass = 3;
+      break;
+    case Hexagon::DuplexIClass4:
+      dupIClass = 4;
+      break;
+    case Hexagon::DuplexIClass5:
+      dupIClass = 5;
+      break;
+    case Hexagon::DuplexIClass6:
+      dupIClass = 6;
+      break;
+    case Hexagon::DuplexIClass7:
+      dupIClass = 7;
+      break;
+    case Hexagon::DuplexIClass8:
+      dupIClass = 8;
+      break;
+    case Hexagon::DuplexIClass9:
+      dupIClass = 9;
+      break;
+    case Hexagon::DuplexIClassA:
+      dupIClass = 10;
+      break;
+    case Hexagon::DuplexIClassB:
+      dupIClass = 11;
+      break;
+    case Hexagon::DuplexIClassC:
+      dupIClass = 12;
+      break;
+    case Hexagon::DuplexIClassD:
+      dupIClass = 13;
+      break;
+    case Hexagon::DuplexIClassE:
+      dupIClass = 14;
+      break;
+    case Hexagon::DuplexIClassF:
+      dupIClass = 15;
+      break;
+    default:
+      llvm_unreachable("Unimplemented DuplexIClass");
+      break;
+    }
+    // 29 is the bit position.
+    // 0b1110 =0xE bits are masked off and down shifted by 1 bit.
+    // Last bit is moved to bit position 13
+    Binary = ((dupIClass & 0xE) << (29 - 1)) | ((dupIClass & 0x1) << 13);
+
+    const MCInst *subInst0 = HMB.getOperand(0).getInst();
+    const MCInst *subInst1 = HMB.getOperand(1).getInst();
+
+    // get subinstruction slot 0
+    unsigned subInstSlot0Bits = getBinaryCodeForInstr(*subInst0, Fixups, STI);
+    // get subinstruction slot 1
+    unsigned subInstSlot1Bits = getBinaryCodeForInstr(*subInst1, Fixups, STI);
+
+    Binary |= subInstSlot0Bits | (subInstSlot1Bits << 16);
+  }
   support::endian::Writer<support::little>(OS).write<uint32_t>(Binary);
   ++MCNumEmitted;
 }
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCDuplexInfo.cpp b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCDuplexInfo.cpp
new file mode 100644
index 00000000000..eb629774a2c
--- /dev/null
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCDuplexInfo.cpp
@@ -0,0 +1,1100 @@
+//===----- HexagonMCDuplexInfo.cpp - Instruction bundle checking ----------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This implements duplexing of instructions to reduce code size
+//
+//===----------------------------------------------------------------------===//
+
+#include "HexagonBaseInfo.h"
+#include "MCTargetDesc/HexagonMCInstrInfo.h"
+
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+
+#include <map>
+
+using namespace llvm;
+using namespace Hexagon;
+
+#define DEBUG_TYPE "hexagon-mcduplex-info"
+
+// pair table of subInstructions with opcodes
+static std::pair<unsigned, unsigned> opcodeData[] = {
+    std::make_pair((unsigned)V4_SA1_addi, 0),
+    std::make_pair((unsigned)V4_SA1_addrx, 6144),
+    std::make_pair((unsigned)V4_SA1_addsp, 3072),
+    std::make_pair((unsigned)V4_SA1_and1, 4608),
+    std::make_pair((unsigned)V4_SA1_clrf, 6768),
+    std::make_pair((unsigned)V4_SA1_clrfnew, 6736),
+    std::make_pair((unsigned)V4_SA1_clrt, 6752),
+    std::make_pair((unsigned)V4_SA1_clrtnew, 6720),
+    std::make_pair((unsigned)V4_SA1_cmpeqi, 6400),
+    std::make_pair((unsigned)V4_SA1_combine0i, 7168),
+    std::make_pair((unsigned)V4_SA1_combine1i, 7176),
+    std::make_pair((unsigned)V4_SA1_combine2i, 7184),
+    std::make_pair((unsigned)V4_SA1_combine3i, 7192),
+    std::make_pair((unsigned)V4_SA1_combinerz, 7432),
+    std::make_pair((unsigned)V4_SA1_combinezr, 7424),
+    std::make_pair((unsigned)V4_SA1_dec, 4864),
+    std::make_pair((unsigned)V4_SA1_inc, 4352),
+    std::make_pair((unsigned)V4_SA1_seti, 2048),
+    std::make_pair((unsigned)V4_SA1_setin1, 6656),
+    std::make_pair((unsigned)V4_SA1_sxtb, 5376),
+    std::make_pair((unsigned)V4_SA1_sxth, 5120),
+    std::make_pair((unsigned)V4_SA1_tfr, 4096),
+    std::make_pair((unsigned)V4_SA1_zxtb, 5888),
+    std::make_pair((unsigned)V4_SA1_zxth, 5632),
+    std::make_pair((unsigned)V4_SL1_loadri_io, 0),
+    std::make_pair((unsigned)V4_SL1_loadrub_io, 4096),
+    std::make_pair((unsigned)V4_SL2_deallocframe, 7936),
+    std::make_pair((unsigned)V4_SL2_jumpr31, 8128),
+    std::make_pair((unsigned)V4_SL2_jumpr31_f, 8133),
+    std::make_pair((unsigned)V4_SL2_jumpr31_fnew, 8135),
+    std::make_pair((unsigned)V4_SL2_jumpr31_t, 8132),
+    std::make_pair((unsigned)V4_SL2_jumpr31_tnew, 8134),
+    std::make_pair((unsigned)V4_SL2_loadrb_io, 4096),
+    std::make_pair((unsigned)V4_SL2_loadrd_sp, 7680),
+    std::make_pair((unsigned)V4_SL2_loadrh_io, 0),
+    std::make_pair((unsigned)V4_SL2_loadri_sp, 7168),
+    std::make_pair((unsigned)V4_SL2_loadruh_io, 2048),
+    std::make_pair((unsigned)V4_SL2_return, 8000),
+    std::make_pair((unsigned)V4_SL2_return_f, 8005),
+    std::make_pair((unsigned)V4_SL2_return_fnew, 8007),
+    std::make_pair((unsigned)V4_SL2_return_t, 8004),
+    std::make_pair((unsigned)V4_SL2_return_tnew, 8006),
+    std::make_pair((unsigned)V4_SS1_storeb_io, 4096),
+    std::make_pair((unsigned)V4_SS1_storew_io, 0),
+    std::make_pair((unsigned)V4_SS2_allocframe, 7168),
+    std::make_pair((unsigned)V4_SS2_storebi0, 4608),
+    std::make_pair((unsigned)V4_SS2_storebi1, 4864),
+    std::make_pair((unsigned)V4_SS2_stored_sp, 2560),
+    std::make_pair((unsigned)V4_SS2_storeh_io, 0),
+    std::make_pair((unsigned)V4_SS2_storew_sp, 2048),
+    std::make_pair((unsigned)V4_SS2_storewi0, 4096),
+    std::make_pair((unsigned)V4_SS2_storewi1, 4352)};
+
+static std::map<unsigned, unsigned>
+    subinstOpcodeMap(opcodeData,
+                     opcodeData + sizeof(opcodeData) / sizeof(opcodeData[0]));
+
+bool HexagonMCInstrInfo::isDuplexPairMatch(unsigned Ga, unsigned Gb) {
+  switch (Ga) {
+  case HexagonII::HSIG_None:
+  default:
+    return false;
+  case HexagonII::HSIG_L1:
+    return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_A);
+  case HexagonII::HSIG_L2:
+    return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 ||
+            Gb == HexagonII::HSIG_A);
+  case HexagonII::HSIG_S1:
+    return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 ||
+            Gb == HexagonII::HSIG_S1 || Gb == HexagonII::HSIG_A);
+  case HexagonII::HSIG_S2:
+    return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 ||
+            Gb == HexagonII::HSIG_S1 || Gb == HexagonII::HSIG_S2 ||
+            Gb == HexagonII::HSIG_A);
+  case HexagonII::HSIG_A:
+    return (Gb == HexagonII::HSIG_A);
+  case HexagonII::HSIG_Compound:
+    return (Gb == HexagonII::HSIG_Compound);
+  }
+  return false;
+}
+
+unsigned HexagonMCInstrInfo::iClassOfDuplexPair(unsigned Ga, unsigned Gb) {
+  switch (Ga) {
+  case HexagonII::HSIG_None:
+  default:
+    break;
+  case HexagonII::HSIG_L1:
+    switch (Gb) {
+    default:
+      break;
+    case HexagonII::HSIG_L1:
+      return 0;
+    case HexagonII::HSIG_A:
+      return 0x4;
+    }
+  case HexagonII::HSIG_L2:
+    switch (Gb) {
+    default:
+      break;
+    case HexagonII::HSIG_L1:
+      return 0x1;
+    case HexagonII::HSIG_L2:
+      return 0x2;
+    case HexagonII::HSIG_A:
+      return 0x5;
+    }
+  case HexagonII::HSIG_S1:
+    switch (Gb) {
+    default:
+      break;
+    case HexagonII::HSIG_L1:
+      return 0x8;
+    case HexagonII::HSIG_L2:
+      return 0x9;
+    case HexagonII::HSIG_S1:
+      return 0xA;
+    case HexagonII::HSIG_A:
+      return 0x6;
+    }
+  case HexagonII::HSIG_S2:
+    switch (Gb) {
+    default:
+      break;
+    case HexagonII::HSIG_L1:
+      return 0xC;
+    case HexagonII::HSIG_L2:
+      return 0xD;
+    case HexagonII::HSIG_S1:
+      return 0xB;
+    case HexagonII::HSIG_S2:
+      return 0xE;
+    case HexagonII::HSIG_A:
+      return 0x7;
+    }
+  case HexagonII::HSIG_A:
+    switch (Gb) {
+    default:
+      break;
+    case HexagonII::HSIG_A:
+      return 0x3;
+    }
+  case HexagonII::HSIG_Compound:
+    switch (Gb) {
+    case HexagonII::HSIG_Compound:
+      return 0xFFFFFFFF;
+    }
+  }
+  return 0xFFFFFFFF;
+}
+
+unsigned HexagonMCInstrInfo::getDuplexCandidateGroup(MCInst const &MCI) {
+  unsigned DstReg, PredReg, SrcReg, Src1Reg, Src2Reg;
+
+  switch (MCI.getOpcode()) {
+  default:
+    return HexagonII::HSIG_None;
+  //
+  // Group L1:
+  //
+  // Rd = memw(Rs+#u4:2)
+  // Rd = memub(Rs+#u4:0)
+  case Hexagon::L2_loadri_io:
+    DstReg = MCI.getOperand(0).getReg();
+    SrcReg = MCI.getOperand(1).getReg();
+    // Special case this one from Group L2.
+    // Rd = memw(r29+#u5:2)
+    if (HexagonMCInstrInfo::isIntRegForSubInst(DstReg)) {
+      if (HexagonMCInstrInfo::isIntReg(SrcReg) && Hexagon::R29 == SrcReg &&
+          MCI.getOperand(2).isImm() &&
+          isShiftedUInt<5, 2>(MCI.getOperand(2).getImm())) {
+        return HexagonII::HSIG_L2;
+      }
+      // Rd = memw(Rs+#u4:2)
+      if (HexagonMCInstrInfo::isIntRegForSubInst(SrcReg) &&
+          (MCI.getOperand(2).isImm() &&
+           isShiftedUInt<4, 2>(MCI.getOperand(2).getImm()))) {
+        return HexagonII::HSIG_L1;
+      }
+    }
+    break;
+  case Hexagon::L2_loadrub_io:
+    // Rd = memub(Rs+#u4:0)
+    DstReg = MCI.getOperand(0).getReg();
+    SrcReg = MCI.getOperand(1).getReg();
+    if (HexagonMCInstrInfo::isIntRegForSubInst(DstReg) &&
+        HexagonMCInstrInfo::isIntRegForSubInst(SrcReg) &&
+        MCI.getOperand(2).isImm() && isUInt<4>(MCI.getOperand(2).getImm())) {
+      return HexagonII::HSIG_L1;
+    }
+    break;
+  //
+  // Group L2:
+  //
+  // Rd = memh/memuh(Rs+#u3:1)
+  // Rd = memb(Rs+#u3:0)
+  // Rd = memw(r29+#u5:2) - Handled above.
+  // Rdd = memd(r29+#u5:3)
+  // deallocframe
+  // [if ([!]p0[.new])] dealloc_return
+  // [if ([!]p0[.new])] jumpr r31
+  case Hexagon::L2_loadrh_io:
+  case Hexagon::L2_loadruh_io:
+    // Rd = memh/memuh(Rs+#u3:1)
+    DstReg = MCI.getOperand(0).getReg();
+    SrcReg = MCI.getOperand(1).getReg();
+    if (HexagonMCInstrInfo::isIntRegForSubInst(DstReg) &&
+        HexagonMCInstrInfo::isIntRegForSubInst(SrcReg) &&
+        MCI.getOperand(2).isImm() &&
+        isShiftedUInt<3, 1>(MCI.getOperand(2).getImm())) {
+      return HexagonII::HSIG_L2;
+    }
+    break;
+  case Hexagon::L2_loadrb_io:
+    // Rd = memb(Rs+#u3:0)
+    DstReg = MCI.getOperand(0).getReg();
+    SrcReg = MCI.getOperand(1).getReg();
+    if (HexagonMCInstrInfo::isIntRegForSubInst(DstReg) &&
+        HexagonMCInstrInfo::isIntRegForSubInst(SrcReg) &&
+        MCI.getOperand(2).isImm() && isUInt<3>(MCI.getOperand(2).getImm())) {
+      return HexagonII::HSIG_L2;
+    }
+    break;
+  case Hexagon::L2_loadrd_io:
+    // Rdd = memd(r29+#u5:3)
+    DstReg = MCI.getOperand(0).getReg();
+    SrcReg = MCI.getOperand(1).getReg();
+    if (HexagonMCInstrInfo::isDblRegForSubInst(DstReg) &&
+        HexagonMCInstrInfo::isIntReg(SrcReg) && Hexagon::R29 == SrcReg &&
+        MCI.getOperand(2).isImm() &&
+        isShiftedUInt<5, 3>(MCI.getOperand(2).getImm())) {
+      return HexagonII::HSIG_L2;
+    }
+    break;
+
+  case Hexagon::L4_return:
+
+  case Hexagon::L2_deallocframe:
+
+    return HexagonII::HSIG_L2;
+  case Hexagon::EH_RETURN_JMPR:
+
+  case Hexagon::J2_jumpr:
+  case Hexagon::JMPret:
+    // jumpr r31
+    // Actual form JMPR %PC<imp-def>, %R31<imp-use>, %R0<imp-use,internal>.
+    DstReg = MCI.getOperand(0).getReg();
+    if (Hexagon::R31 == DstReg) {
+      return HexagonII::HSIG_L2;
+    }
+    break;
+
+  case Hexagon::J2_jumprt:
+  case Hexagon::J2_jumprf:
+  case Hexagon::J2_jumprtnew:
+  case Hexagon::J2_jumprfnew:
+  case Hexagon::JMPrett:
+  case Hexagon::JMPretf:
+  case Hexagon::JMPrettnew:
+  case Hexagon::JMPretfnew:
+  case Hexagon::JMPrettnewpt:
+  case Hexagon::JMPretfnewpt:
+    DstReg = MCI.getOperand(1).getReg();
+    SrcReg = MCI.getOperand(0).getReg();
+    // [if ([!]p0[.new])] jumpr r31
+    if ((HexagonMCInstrInfo::isPredReg(SrcReg) && (Hexagon::P0 == SrcReg)) &&
+        (Hexagon::R31 == DstReg)) {
+      return HexagonII::HSIG_L2;
+    }
+    break;
+  case Hexagon::L4_return_t:
+
+  case Hexagon::L4_return_f:
+
+  case Hexagon::L4_return_tnew_pnt:
+
+  case Hexagon::L4_return_fnew_pnt:
+
+  case Hexagon::L4_return_tnew_pt:
+
+  case Hexagon::L4_return_fnew_pt:
+    // [if ([!]p0[.new])] dealloc_return
+    SrcReg = MCI.getOperand(0).getReg();
+    if (Hexagon::P0 == SrcReg) {
+      return HexagonII::HSIG_L2;
+    }
+    break;
+  //
+  // Group S1:
+  //
+  // memw(Rs+#u4:2) = Rt
+  // memb(Rs+#u4:0) = Rt
+  case Hexagon::S2_storeri_io:
+    // Special case this one from Group S2.
+    // memw(r29+#u5:2) = Rt
+    Src1Reg = MCI.getOperand(0).getReg();
+    Src2Reg = MCI.getOperand(2).getReg();
+    if (HexagonMCInstrInfo::isIntReg(Src1Reg) &&
+        HexagonMCInstrInfo::isIntRegForSubInst(Src2Reg) &&
+        Hexagon::R29 == Src1Reg && MCI.getOperand(1).isImm() &&
+        isShiftedUInt<5, 2>(MCI.getOperand(1).getImm())) {
+      return HexagonII::HSIG_S2;
+    }
+    // memw(Rs+#u4:2) = Rt
+    if (HexagonMCInstrInfo::isIntRegForSubInst(Src1Reg) &&
+        HexagonMCInstrInfo::isIntRegForSubInst(Src2Reg) &&
+        MCI.getOperand(1).isImm() &&
+        isShiftedUInt<4, 2>(MCI.getOperand(1).getImm())) {
+      return HexagonII::HSIG_S1;
+    }
+    break;
+  case Hexagon::S2_storerb_io:
+    // memb(Rs+#u4:0) = Rt
+    Src1Reg = MCI.getOperand(0).getReg();
+    Src2Reg = MCI.getOperand(2).getReg();
+    if (HexagonMCInstrInfo::isIntRegForSubInst(Src1Reg) &&
+        HexagonMCInstrInfo::isIntRegForSubInst(Src2Reg) &&
+        MCI.getOperand(1).isImm() && isUInt<4>(MCI.getOperand(1).getImm())) {
+      return HexagonII::HSIG_S1;
+    }
+    break;
+  //
+  // Group S2:
+  //
+  // memh(Rs+#u3:1) = Rt
+  // memw(r29+#u5:2) = Rt
+  // memd(r29+#s6:3) = Rtt
+  // memw(Rs+#u4:2) = #U1
+  // memb(Rs+#u4) = #U1
+  // allocframe(#u5:3)
+  case Hexagon::S2_storerh_io:
+    // memh(Rs+#u3:1) = Rt
+    Src1Reg = MCI.getOperand(0).getReg();
+    Src2Reg = MCI.getOperand(2).getReg();
+    if (HexagonMCInstrInfo::isIntRegForSubInst(Src1Reg) &&
+        HexagonMCInstrInfo::isIntRegForSubInst(Src2Reg) &&
+        MCI.getOperand(1).isImm() &&
+        isShiftedUInt<3, 1>(MCI.getOperand(1).getImm())) {
+      return HexagonII::HSIG_S2;
+    }
+    break;
+  case Hexagon::S2_storerd_io:
+    // memd(r29+#s6:3) = Rtt
+    Src1Reg = MCI.getOperand(0).getReg();
+    Src2Reg = MCI.getOperand(2).getReg();
+    if (HexagonMCInstrInfo::isDblRegForSubInst(Src2Reg) &&
+        HexagonMCInstrInfo::isIntReg(Src1Reg) && Hexagon::R29 == Src1Reg &&
+        MCI.getOperand(1).isImm() &&
+        isShiftedInt<6, 3>(MCI.getOperand(1).getImm())) {
+      return HexagonII::HSIG_S2;
+    }
+    break;
+  case Hexagon::S4_storeiri_io:
+    // memw(Rs+#u4:2) = #U1
+    Src1Reg = MCI.getOperand(0).getReg();
+    if (HexagonMCInstrInfo::isIntRegForSubInst(Src1Reg) &&
+        MCI.getOperand(1).isImm() &&
+        isShiftedUInt<4, 2>(MCI.getOperand(1).getImm()) &&
+        MCI.getOperand(2).isImm() && isUInt<1>(MCI.getOperand(2).getImm())) {
+      return HexagonII::HSIG_S2;
+    }
+    break;
+  case Hexagon::S4_storeirb_io:
+    // memb(Rs+#u4) = #U1
+    Src1Reg = MCI.getOperand(0).getReg();
+    if (HexagonMCInstrInfo::isIntRegForSubInst(Src1Reg) &&
+        MCI.getOperand(1).isImm() && isUInt<4>(MCI.getOperand(1).getImm()) &&
+        MCI.getOperand(2).isImm() && MCI.getOperand(2).isImm() &&
+        isUInt<1>(MCI.getOperand(2).getImm())) {
+      return HexagonII::HSIG_S2;
+    }
+    break;
+  case Hexagon::S2_allocframe:
+    if (MCI.getOperand(0).isImm() &&
+        isShiftedUInt<5, 3>(MCI.getOperand(0).getImm())) {
+      return HexagonII::HSIG_S2;
+    }
+    break;
+  //
+  // Group A:
+  //
+  // Rx = add(Rx,#s7)
+  // Rd = Rs
+  // Rd = #u6
+  // Rd = #-1
+  // if ([!]P0[.new]) Rd = #0
+  // Rd = add(r29,#u6:2)
+  // Rx = add(Rx,Rs)
+  // P0 = cmp.eq(Rs,#u2)
+  // Rdd = combine(#0,Rs)
+  // Rdd = combine(Rs,#0)
+  // Rdd = combine(#u2,#U2)
+  // Rd = add(Rs,#1)
+  // Rd = add(Rs,#-1)
+  // Rd = sxth/sxtb/zxtb/zxth(Rs)
+  // Rd = and(Rs,#1)
+  case Hexagon::A2_addi:
+    DstReg = MCI.getOperand(0).getReg();
+    SrcReg = MCI.getOperand(1).getReg();
+    if (HexagonMCInstrInfo::isIntRegForSubInst(DstReg)) {
+      // Rd = add(r29,#u6:2)
+      if (HexagonMCInstrInfo::isIntReg(SrcReg) && Hexagon::R29 == SrcReg &&
+          MCI.getOperand(2).isImm() &&
+          isShiftedUInt<6, 2>(MCI.getOperand(2).getImm())) {
+        return HexagonII::HSIG_A;
+      }
+      // Rx = add(Rx,#s7)
+      if (DstReg == SrcReg) {
+        return HexagonII::HSIG_A;
+      }
+      // Rd = add(Rs,#1)
+      // Rd = add(Rs,#-1)
+      if (HexagonMCInstrInfo::isIntRegForSubInst(SrcReg) &&
+          MCI.getOperand(2).isImm() && ((MCI.getOperand(2).getImm() == 1) ||
+                                        (MCI.getOperand(2).getImm() == -1))) {
+        return HexagonII::HSIG_A;
+      }
+    }
+    break;
+  case Hexagon::A2_add:
+    // Rx = add(Rx,Rs)
+    DstReg = MCI.getOperand(0).getReg();
+    Src1Reg = MCI.getOperand(1).getReg();
+    Src2Reg = MCI.getOperand(2).getReg();
+    if (HexagonMCInstrInfo::isIntRegForSubInst(DstReg) && (DstReg == Src1Reg) &&
+        HexagonMCInstrInfo::isIntRegForSubInst(Src2Reg)) {
+      return HexagonII::HSIG_A;
+    }
+    break;
+  case Hexagon::A2_andir:
+    DstReg = MCI.getOperand(0).getReg();
+    SrcReg = MCI.getOperand(1).getReg();
+    if (HexagonMCInstrInfo::isIntRegForSubInst(DstReg) &&
+        HexagonMCInstrInfo::isIntRegForSubInst(SrcReg) &&
+        MCI.getOperand(2).isImm() && ((MCI.getOperand(2).getImm() == 1) ||
+                                      (MCI.getOperand(2).getImm() == 255))) {
+      return HexagonII::HSIG_A;
+    }
+    break;
+  case Hexagon::A2_tfr:
+    // Rd = Rs
+    DstReg = MCI.getOperand(0).getReg();
+    SrcReg = MCI.getOperand(1).getReg();
+    if (HexagonMCInstrInfo::isIntRegForSubInst(DstReg) &&
+        HexagonMCInstrInfo::isIntRegForSubInst(SrcReg)) {
+      return HexagonII::HSIG_A;
+    }
+    break;
+  case Hexagon::A2_tfrsi:
+    DstReg = MCI.getOperand(0).getReg();
+
+    if (HexagonMCInstrInfo::isIntRegForSubInst(DstReg)) {
+      return HexagonII::HSIG_A;
+    }
+    break;
+  case Hexagon::C2_cmoveit:
+  case Hexagon::C2_cmovenewit:
+  case Hexagon::C2_cmoveif:
+  case Hexagon::C2_cmovenewif:
+    // if ([!]P0[.new]) Rd = #0
+    // Actual form:
+    // %R16<def> = C2_cmovenewit %P0<internal>, 0, %R16<imp-use,undef>;
+    DstReg = MCI.getOperand(0).getReg();  // Rd
+    PredReg = MCI.getOperand(1).getReg(); // P0
+    if (HexagonMCInstrInfo::isIntRegForSubInst(DstReg) &&
+        Hexagon::P0 == PredReg && MCI.getOperand(2).isImm() &&
+        MCI.getOperand(2).getImm() == 0) {
+      return HexagonII::HSIG_A;
+    }
+    break;
+  case Hexagon::C2_cmpeqi:
+    // P0 = cmp.eq(Rs,#u2)
+    DstReg = MCI.getOperand(0).getReg();
+    SrcReg = MCI.getOperand(1).getReg();
+    if (Hexagon::P0 == DstReg &&
+        HexagonMCInstrInfo::isIntRegForSubInst(SrcReg) &&
+        MCI.getOperand(2).isImm() && isUInt<2>(MCI.getOperand(2).getImm())) {
+      return HexagonII::HSIG_A;
+    }
+    break;
+  case Hexagon::A2_combineii:
+  case Hexagon::A4_combineii:
+    // Rdd = combine(#u2,#U2)
+    DstReg = MCI.getOperand(0).getReg();
+    if (HexagonMCInstrInfo::isDblRegForSubInst(DstReg) &&
+        // TODO: Handle Globals/Symbols
+        (MCI.getOperand(1).isImm() && isUInt<2>(MCI.getOperand(1).getImm())) &&
+        ((MCI.getOperand(2).isImm() &&
+          isUInt<2>(MCI.getOperand(2).getImm())))) {
+      return HexagonII::HSIG_A;
+    }
+    break;
+  case Hexagon::A4_combineri:
+    // Rdd = combine(Rs,#0)
+    DstReg = MCI.getOperand(0).getReg();
+    SrcReg = MCI.getOperand(1).getReg();
+    if (HexagonMCInstrInfo::isDblRegForSubInst(DstReg) &&
+        HexagonMCInstrInfo::isIntRegForSubInst(SrcReg) &&
+        (MCI.getOperand(2).isImm() && MCI.getOperand(2).getImm() == 0)) {
+      return HexagonII::HSIG_A;
+    }
+    break;
+  case Hexagon::A4_combineir:
+    // Rdd = combine(#0,Rs)
+    DstReg = MCI.getOperand(0).getReg();
+    SrcReg = MCI.getOperand(2).getReg();
+    if (HexagonMCInstrInfo::isDblRegForSubInst(DstReg) &&
+        HexagonMCInstrInfo::isIntRegForSubInst(SrcReg) &&
+        (MCI.getOperand(1).isImm() && MCI.getOperand(1).getImm() == 0)) {
+      return HexagonII::HSIG_A;
+    }
+    break;
+  case Hexagon::A2_sxtb:
+  case Hexagon::A2_sxth:
+  case Hexagon::A2_zxtb:
+  case Hexagon::A2_zxth:
+    // Rd = sxth/sxtb/zxtb/zxth(Rs)
+    DstReg = MCI.getOperand(0).getReg();
+    SrcReg = MCI.getOperand(1).getReg();
+    if (HexagonMCInstrInfo::isIntRegForSubInst(DstReg) &&
+        HexagonMCInstrInfo::isIntRegForSubInst(SrcReg)) {
+      return HexagonII::HSIG_A;
+    }
+    break;
+  }
+
+  return HexagonII::HSIG_None;
+}
+
+bool HexagonMCInstrInfo::subInstWouldBeExtended(MCInst const &potentialDuplex) {
+
+  unsigned DstReg, SrcReg;
+
+  switch (potentialDuplex.getOpcode()) {
+  case Hexagon::A2_addi:
+    // testing for case of: Rx = add(Rx,#s7)
+    DstReg = potentialDuplex.getOperand(0).getReg();
+    SrcReg = potentialDuplex.getOperand(1).getReg();
+    if (DstReg == SrcReg && HexagonMCInstrInfo::isIntRegForSubInst(DstReg)) {
+      if (potentialDuplex.getOperand(2).isExpr())
+        return true;
+      if (potentialDuplex.getOperand(2).isImm() &&
+          !(isShiftedInt<7, 0>(potentialDuplex.getOperand(2).getImm())))
+        return true;
+    }
+    break;
+  case Hexagon::A2_tfrsi:
+    DstReg = potentialDuplex.getOperand(0).getReg();
+
+    if (HexagonMCInstrInfo::isIntRegForSubInst(DstReg)) {
+      if (potentialDuplex.getOperand(1).isExpr())
+        return true;
+      // Check for case of Rd = #-1.
+      if (potentialDuplex.getOperand(1).isImm() &&
+          (potentialDuplex.getOperand(1).getImm() == -1))
+        return false;
+      // Check for case of Rd = #u6.
+      if (potentialDuplex.getOperand(1).isImm() &&
+          !isShiftedUInt<6, 0>(potentialDuplex.getOperand(1).getImm()))
+        return true;
+    }
+    break;
+  default:
+    break;
+  }
+  return false;
+}
+
+/// non-Symmetrical. See if these two instructions are fit for duplex pair.
+bool HexagonMCInstrInfo::isOrderedDuplexPair(MCInstrInfo const &MCII,
+                                             MCInst const &MIa, bool ExtendedA,
+                                             MCInst const &MIb, bool ExtendedB,
+                                             bool bisReversable) {
+  // Slot 1 cannot be extended in duplexes PRM 10.5
+  if (ExtendedA)
+    return false;
+  // Only A2_addi and A2_tfrsi can be extended in duplex form PRM 10.5
+  if (ExtendedB) {
+    unsigned Opcode = MIb.getOpcode();
+    if ((Opcode != Hexagon::A2_addi) && (Opcode != Hexagon::A2_tfrsi))
+      return false;
+  }
+  unsigned MIaG = HexagonMCInstrInfo::getDuplexCandidateGroup(MIa),
+           MIbG = HexagonMCInstrInfo::getDuplexCandidateGroup(MIb);
+
+  // If a duplex contains 2 insns in the same group, the insns must be
+  // ordered such that the numerically smaller opcode is in slot 1.
+  if ((MIaG != HexagonII::HSIG_None) && (MIaG == MIbG) && bisReversable) {
+    MCInst SubInst0 = HexagonMCInstrInfo::deriveSubInst(MIa);
+    MCInst SubInst1 = HexagonMCInstrInfo::deriveSubInst(MIb);
+
+    unsigned zeroedSubInstS0 =
+        subinstOpcodeMap.find(SubInst0.getOpcode())->second;
+    unsigned zeroedSubInstS1 =
+        subinstOpcodeMap.find(SubInst1.getOpcode())->second;
+
+    if (zeroedSubInstS0 < zeroedSubInstS1)
+      // subinstS0 (maps to slot 0) must be greater than
+      // subinstS1 (maps to slot 1)
+      return false;
+  }
+
+  // allocframe must always be in slot 0
+  if (MIb.getOpcode() == Hexagon::S2_allocframe)
+    return false;
+
+  if ((MIaG != HexagonII::HSIG_None) && (MIbG != HexagonII::HSIG_None)) {
+    // Prevent 2 instructions with extenders from duplexing
+    // Note that MIb (slot1) can be extended and MIa (slot0)
+    //   can never be extended
+    if (subInstWouldBeExtended(MIa))
+      return false;
+
+    // If duplexing produces an extender, but the original did not
+    //   have an extender, do not duplex.
+    if (subInstWouldBeExtended(MIb) && !ExtendedB)
+      return false;
+  }
+
+  // If jumpr r31 appears, it must be in slot 0, and never slot 1 (MIb).
+  if (MIbG == HexagonII::HSIG_L2) {
+    if ((MIb.getNumOperands() > 1) && MIb.getOperand(1).isReg() &&
+        (MIb.getOperand(1).getReg() == Hexagon::R31))
+      return false;
+    if ((MIb.getNumOperands() > 0) && MIb.getOperand(0).isReg() &&
+        (MIb.getOperand(0).getReg() == Hexagon::R31))
+      return false;
+  }
+
+  // If a store appears, it must be in slot 0 (MIa) 1st, and then slot 1 (MIb);
+  //   therefore, not duplexable if slot 1 is a store, and slot 0 is not.
+  if ((MIbG == HexagonII::HSIG_S1) || (MIbG == HexagonII::HSIG_S2)) {
+    if ((MIaG != HexagonII::HSIG_S1) && (MIaG != HexagonII::HSIG_S2))
+      return false;
+  }
+
+  return (isDuplexPairMatch(MIaG, MIbG));
+}
+
+/// Symmetrical. See if these two instructions are fit for duplex pair.
+bool HexagonMCInstrInfo::isDuplexPair(MCInst const &MIa, MCInst const &MIb) {
+  unsigned MIaG = getDuplexCandidateGroup(MIa),
+           MIbG = getDuplexCandidateGroup(MIb);
+  return (isDuplexPairMatch(MIaG, MIbG) || isDuplexPairMatch(MIbG, MIaG));
+}
+
+inline static void addOps(MCInst &subInstPtr, MCInst const &Inst,
+                          unsigned opNum) {
+  if (Inst.getOperand(opNum).isReg()) {
+    switch (Inst.getOperand(opNum).getReg()) {
+    default:
+      llvm_unreachable("Not Duplexable Register");
+      break;
+    case Hexagon::R0:
+    case Hexagon::R1:
+    case Hexagon::R2:
+    case Hexagon::R3:
+    case Hexagon::R4:
+    case Hexagon::R5:
+    case Hexagon::R6:
+    case Hexagon::R7:
+    case Hexagon::D0:
+    case Hexagon::D1:
+    case Hexagon::D2:
+    case Hexagon::D3:
+    case Hexagon::R16:
+    case Hexagon::R17:
+    case Hexagon::R18:
+    case Hexagon::R19:
+    case Hexagon::R20:
+    case Hexagon::R21:
+    case Hexagon::R22:
+    case Hexagon::R23:
+    case Hexagon::D8:
+    case Hexagon::D9:
+    case Hexagon::D10:
+    case Hexagon::D11:
+      subInstPtr.addOperand(Inst.getOperand(opNum));
+      break;
+    }
+  } else
+    subInstPtr.addOperand(Inst.getOperand(opNum));
+}
+
+MCInst HexagonMCInstrInfo::deriveSubInst(MCInst const &Inst) {
+  MCInst Result;
+  switch (Inst.getOpcode()) {
+  default:
+    // dbgs() << "opcode: "<< Inst->getOpcode() << "\n";
+    llvm_unreachable("Unimplemented subinstruction \n");
+    break;
+  case Hexagon::A2_addi:
+    if (Inst.getOperand(2).isImm() && Inst.getOperand(2).getImm() == 1) {
+      Result.setOpcode(Hexagon::V4_SA1_inc);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 1);
+      break;
+    } //  1,2 SUBInst $Rd = add($Rs, #1)
+    else if (Inst.getOperand(2).isImm() && Inst.getOperand(2).getImm() == -1) {
+      Result.setOpcode(Hexagon::V4_SA1_dec);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 1);
+      break;
+    } //  1,2 SUBInst $Rd = add($Rs,#-1)
+    else if (Inst.getOperand(1).getReg() == Hexagon::R29) {
+      Result.setOpcode(Hexagon::V4_SA1_addsp);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 2);
+      break;
+    } //  1,3 SUBInst $Rd = add(r29, #$u6_2)
+    else {
+      Result.setOpcode(Hexagon::V4_SA1_addi);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 1);
+      addOps(Result, Inst, 2);
+      break;
+    } //    1,2,3 SUBInst $Rx = add($Rx, #$s7)
+  case Hexagon::A2_add:
+    Result.setOpcode(Hexagon::V4_SA1_addrx);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 1);
+    addOps(Result, Inst, 2);
+    break; //    1,2,3 SUBInst $Rx = add($_src_, $Rs)
+  case Hexagon::S2_allocframe:
+    Result.setOpcode(Hexagon::V4_SS2_allocframe);
+    addOps(Result, Inst, 0);
+    break; //    1 SUBInst allocframe(#$u5_3)
+  case Hexagon::A2_andir:
+    if (Inst.getOperand(2).getImm() == 255) {
+      Result.setOpcode(Hexagon::V4_SA1_zxtb);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 1);
+      break; //    1,2    $Rd = and($Rs, #255)
+    } else {
+      Result.setOpcode(Hexagon::V4_SA1_and1);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 1);
+      break; //    1,2 SUBInst $Rd = and($Rs, #1)
+    }
+  case Hexagon::C2_cmpeqi:
+    Result.setOpcode(Hexagon::V4_SA1_cmpeqi);
+    addOps(Result, Inst, 1);
+    addOps(Result, Inst, 2);
+    break; //    2,3 SUBInst p0 = cmp.eq($Rs, #$u2)
+  case Hexagon::A4_combineii:
+  case Hexagon::A2_combineii:
+    if (Inst.getOperand(1).getImm() == 1) {
+      Result.setOpcode(Hexagon::V4_SA1_combine1i);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 2);
+      break; //  1,3 SUBInst $Rdd = combine(#1, #$u2)
+    }
+
+    if (Inst.getOperand(1).getImm() == 3) {
+      Result.setOpcode(Hexagon::V4_SA1_combine3i);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 2);
+      break; //  1,3 SUBInst $Rdd = combine(#3, #$u2)
+    }
+    if (Inst.getOperand(1).getImm() == 0) {
+      Result.setOpcode(Hexagon::V4_SA1_combine0i);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 2);
+      break; //  1,3 SUBInst $Rdd = combine(#0, #$u2)
+    }
+    if (Inst.getOperand(1).getImm() == 2) {
+      Result.setOpcode(Hexagon::V4_SA1_combine2i);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 2);
+      break; //  1,3 SUBInst $Rdd = combine(#2, #$u2)
+    }
+  case Hexagon::A4_combineir:
+    Result.setOpcode(Hexagon::V4_SA1_combinezr);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 2);
+    break; //    1,3 SUBInst $Rdd = combine(#0, $Rs)
+
+  case Hexagon::A4_combineri:
+    Result.setOpcode(Hexagon::V4_SA1_combinerz);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 1);
+    break; //    1,2 SUBInst $Rdd = combine($Rs, #0)
+  case Hexagon::L4_return_tnew_pnt:
+  case Hexagon::L4_return_tnew_pt:
+    Result.setOpcode(Hexagon::V4_SL2_return_tnew);
+    break; //    none  SUBInst if (p0.new) dealloc_return:nt
+  case Hexagon::L4_return_fnew_pnt:
+  case Hexagon::L4_return_fnew_pt:
+    Result.setOpcode(Hexagon::V4_SL2_return_fnew);
+    break; //    none  SUBInst if (!p0.new) dealloc_return:nt
+  case Hexagon::L4_return_f:
+    Result.setOpcode(Hexagon::V4_SL2_return_f);
+    break; //    none  SUBInst if (!p0) dealloc_return
+  case Hexagon::L4_return_t:
+    Result.setOpcode(Hexagon::V4_SL2_return_t);
+    break; //    none  SUBInst if (p0) dealloc_return
+  case Hexagon::L4_return:
+    Result.setOpcode(Hexagon::V4_SL2_return);
+    break; //    none  SUBInst dealloc_return
+  case Hexagon::L2_deallocframe:
+    Result.setOpcode(Hexagon::V4_SL2_deallocframe);
+    break; //    none  SUBInst deallocframe
+  case Hexagon::EH_RETURN_JMPR:
+  case Hexagon::J2_jumpr:
+  case Hexagon::JMPret:
+    Result.setOpcode(Hexagon::V4_SL2_jumpr31);
+    break; //    none  SUBInst jumpr r31
+  case Hexagon::J2_jumprf:
+  case Hexagon::JMPretf:
+    Result.setOpcode(Hexagon::V4_SL2_jumpr31_f);
+    break; //    none  SUBInst if (!p0) jumpr r31
+  case Hexagon::J2_jumprfnew:
+  case Hexagon::JMPretfnewpt:
+  case Hexagon::JMPretfnew:
+    Result.setOpcode(Hexagon::V4_SL2_jumpr31_fnew);
+    break; //    none  SUBInst if (!p0.new) jumpr:nt r31
+  case Hexagon::J2_jumprt:
+  case Hexagon::JMPrett:
+    Result.setOpcode(Hexagon::V4_SL2_jumpr31_t);
+    break; //    none  SUBInst if (p0) jumpr r31
+  case Hexagon::J2_jumprtnew:
+  case Hexagon::JMPrettnewpt:
+  case Hexagon::JMPrettnew:
+    Result.setOpcode(Hexagon::V4_SL2_jumpr31_tnew);
+    break; //    none  SUBInst if (p0.new) jumpr:nt r31
+  case Hexagon::L2_loadrb_io:
+    Result.setOpcode(Hexagon::V4_SL2_loadrb_io);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 1);
+    addOps(Result, Inst, 2);
+    break; //    1,2,3 SUBInst $Rd = memb($Rs + #$u3_0)
+  case Hexagon::L2_loadrd_io:
+    Result.setOpcode(Hexagon::V4_SL2_loadrd_sp);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 2);
+    break; //    1,3 SUBInst $Rdd = memd(r29 + #$u5_3)
+  case Hexagon::L2_loadrh_io:
+    Result.setOpcode(Hexagon::V4_SL2_loadrh_io);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 1);
+    addOps(Result, Inst, 2);
+    break; //    1,2,3 SUBInst $Rd = memh($Rs + #$u3_1)
+  case Hexagon::L2_loadrub_io:
+    Result.setOpcode(Hexagon::V4_SL1_loadrub_io);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 1);
+    addOps(Result, Inst, 2);
+    break; //    1,2,3 SUBInst $Rd = memub($Rs + #$u4_0)
+  case Hexagon::L2_loadruh_io:
+    Result.setOpcode(Hexagon::V4_SL2_loadruh_io);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 1);
+    addOps(Result, Inst, 2);
+    break; //    1,2,3 SUBInst $Rd = memuh($Rs + #$u3_1)
+  case Hexagon::L2_loadri_io:
+    if (Inst.getOperand(1).getReg() == Hexagon::R29) {
+      Result.setOpcode(Hexagon::V4_SL2_loadri_sp);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 2);
+      break; //  2 1,3 SUBInst $Rd = memw(r29 + #$u5_2)
+    } else {
+      Result.setOpcode(Hexagon::V4_SL1_loadri_io);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 1);
+      addOps(Result, Inst, 2);
+      break; //    1,2,3 SUBInst $Rd = memw($Rs + #$u4_2)
+    }
+  case Hexagon::S4_storeirb_io:
+    if (Inst.getOperand(2).getImm() == 0) {
+      Result.setOpcode(Hexagon::V4_SS2_storebi0);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 1);
+      break; //    1,2 SUBInst memb($Rs + #$u4_0)=#0
+    } else if (Inst.getOperand(2).getImm() == 1) {
+      Result.setOpcode(Hexagon::V4_SS2_storebi1);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 1);
+      break; //  2 1,2 SUBInst memb($Rs + #$u4_0)=#1
+    }
+  case Hexagon::S2_storerb_io:
+    Result.setOpcode(Hexagon::V4_SS1_storeb_io);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 1);
+    addOps(Result, Inst, 2);
+    break; //    1,2,3 SUBInst memb($Rs + #$u4_0) = $Rt
+  case Hexagon::S2_storerd_io:
+    Result.setOpcode(Hexagon::V4_SS2_stored_sp);
+    addOps(Result, Inst, 1);
+    addOps(Result, Inst, 2);
+    break; //    2,3 SUBInst memd(r29 + #$s6_3) = $Rtt
+  case Hexagon::S2_storerh_io:
+    Result.setOpcode(Hexagon::V4_SS2_storeh_io);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 1);
+    addOps(Result, Inst, 2);
+    break; //    1,2,3 SUBInst memb($Rs + #$u4_0) = $Rt
+  case Hexagon::S4_storeiri_io:
+    if (Inst.getOperand(2).getImm() == 0) {
+      Result.setOpcode(Hexagon::V4_SS2_storewi0);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 1);
+      break; //  3 1,2 SUBInst memw($Rs + #$u4_2)=#0
+    } else if (Inst.getOperand(2).getImm() == 1) {
+      Result.setOpcode(Hexagon::V4_SS2_storewi1);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 1);
+      break; //  3 1,2 SUBInst memw($Rs + #$u4_2)=#1
+    } else if (Inst.getOperand(0).getReg() == Hexagon::R29) {
+      Result.setOpcode(Hexagon::V4_SS2_storew_sp);
+      addOps(Result, Inst, 1);
+      addOps(Result, Inst, 2);
+      break; //  1 2,3 SUBInst memw(r29 + #$u5_2) = $Rt
+    }
+  case Hexagon::S2_storeri_io:
+    if (Inst.getOperand(0).getReg() == Hexagon::R29) {
+      Result.setOpcode(Hexagon::V4_SS2_storew_sp);
+      addOps(Result, Inst, 1);
+      addOps(Result, Inst, 2); //  1,2,3 SUBInst memw(sp + #$u5_2) = $Rt
+    } else {
+      Result.setOpcode(Hexagon::V4_SS1_storew_io);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 1);
+      addOps(Result, Inst, 2); //  1,2,3 SUBInst memw($Rs + #$u4_2) = $Rt
+    }
+    break;
+  case Hexagon::A2_sxtb:
+    Result.setOpcode(Hexagon::V4_SA1_sxtb);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 1);
+    break; //  1,2 SUBInst $Rd = sxtb($Rs)
+  case Hexagon::A2_sxth:
+    Result.setOpcode(Hexagon::V4_SA1_sxth);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 1);
+    break; //  1,2 SUBInst $Rd = sxth($Rs)
+  case Hexagon::A2_tfr:
+    Result.setOpcode(Hexagon::V4_SA1_tfr);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 1);
+    break; //  1,2 SUBInst $Rd = $Rs
+  case Hexagon::C2_cmovenewif:
+    Result.setOpcode(Hexagon::V4_SA1_clrfnew);
+    addOps(Result, Inst, 0);
+    break; //  2 SUBInst if (!p0.new) $Rd = #0
+  case Hexagon::C2_cmovenewit:
+    Result.setOpcode(Hexagon::V4_SA1_clrtnew);
+    addOps(Result, Inst, 0);
+    break; //  2 SUBInst if (p0.new) $Rd = #0
+  case Hexagon::C2_cmoveif:
+    Result.setOpcode(Hexagon::V4_SA1_clrf);
+    addOps(Result, Inst, 0);
+    break; //  2 SUBInst if (!p0) $Rd = #0
+  case Hexagon::C2_cmoveit:
+    Result.setOpcode(Hexagon::V4_SA1_clrt);
+    addOps(Result, Inst, 0);
+    break; //  2 SUBInst if (p0) $Rd = #0
+  case Hexagon::A2_tfrsi:
+    if (Inst.getOperand(1).isImm() && Inst.getOperand(1).getImm() == -1) {
+      Result.setOpcode(Hexagon::V4_SA1_setin1);
+      addOps(Result, Inst, 0);
+      break; //  2 1 SUBInst $Rd = #-1
+    } else {
+      Result.setOpcode(Hexagon::V4_SA1_seti);
+      addOps(Result, Inst, 0);
+      addOps(Result, Inst, 1);
+      break; //    1,2 SUBInst $Rd = #$u6
+    }
+  case Hexagon::A2_zxtb:
+    Result.setOpcode(Hexagon::V4_SA1_zxtb);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 1);
+    break; //    1,2    $Rd = and($Rs, #255)
+
+  case Hexagon::A2_zxth:
+    Result.setOpcode(Hexagon::V4_SA1_zxth);
+    addOps(Result, Inst, 0);
+    addOps(Result, Inst, 1);
+    break; //    1,2 SUBInst $Rd = zxth($Rs)
+  }
+  return Result;
+}
+
+static bool isStoreInst(unsigned opCode) {
+  switch (opCode) {
+  case Hexagon::S2_storeri_io:
+  case Hexagon::S2_storerb_io:
+  case Hexagon::S2_storerh_io:
+  case Hexagon::S2_storerd_io:
+  case Hexagon::S4_storeiri_io:
+  case Hexagon::S4_storeirb_io:
+  case Hexagon::S2_allocframe:
+    return true;
+  default:
+    return false;
+  }
+}
+
+SmallVector<DuplexCandidate, 8>
+HexagonMCInstrInfo::getDuplexPossibilties(MCInstrInfo const &MCII,
+                                          MCInst const &MCB) {
+  assert(isBundle(MCB));
+  SmallVector<DuplexCandidate, 8> duplexToTry;
+  // Use an "order matters" version of isDuplexPair.
+  unsigned numInstrInPacket = MCB.getNumOperands();
+
+  for (unsigned distance = 1; distance < numInstrInPacket; ++distance) {
+    for (unsigned j = HexagonMCInstrInfo::bundleInstructionsOffset,
+                  k = j + distance;
+         (j < numInstrInPacket) && (k < numInstrInPacket); ++j, ++k) {
+
+      // Check if reversable.
+      bool bisReversable = true;
+      if (isStoreInst(MCB.getOperand(j).getInst()->getOpcode()) &&
+          isStoreInst(MCB.getOperand(k).getInst()->getOpcode())) {
+        DEBUG(dbgs() << "skip out of order write pair: " << k << "," << j
+                     << "\n");
+        bisReversable = false;
+      }
+
+      // Try in order.
+      if (isOrderedDuplexPair(
+              MCII, *MCB.getOperand(k).getInst(),
+              HexagonMCInstrInfo::hasExtenderForIndex(MCB, k - 1),
+              *MCB.getOperand(j).getInst(),
+              HexagonMCInstrInfo::hasExtenderForIndex(MCB, j - 1),
+              bisReversable)) {
+        // Get iClass.
+        unsigned iClass = iClassOfDuplexPair(
+            getDuplexCandidateGroup(*MCB.getOperand(k).getInst()),
+            getDuplexCandidateGroup(*MCB.getOperand(j).getInst()));
+
+        // Save off pairs for duplex checking.
+        duplexToTry.push_back(DuplexCandidate(j, k, iClass));
+        DEBUG(dbgs() << "adding pair: " << j << "," << k << ":"
+                     << MCB.getOperand(j).getInst()->getOpcode() << ","
+                     << MCB.getOperand(k).getInst()->getOpcode() << "\n");
+        continue;
+      } else {
+        DEBUG(dbgs() << "skipping pair: " << j << "," << k << ":"
+                     << MCB.getOperand(j).getInst()->getOpcode() << ","
+                     << MCB.getOperand(k).getInst()->getOpcode() << "\n");
+      }
+
+      // Try reverse.
+      if (bisReversable) {
+        if (isOrderedDuplexPair(
+                MCII, *MCB.getOperand(j).getInst(),
+                HexagonMCInstrInfo::hasExtenderForIndex(MCB, j - 1),
+                *MCB.getOperand(k).getInst(),
+                HexagonMCInstrInfo::hasExtenderForIndex(MCB, k - 1),
+                bisReversable)) {
+          // Get iClass.
+          unsigned iClass = iClassOfDuplexPair(
+              getDuplexCandidateGroup(*MCB.getOperand(j).getInst()),
+              getDuplexCandidateGroup(*MCB.getOperand(k).getInst()));
+
+          // Save off pairs for duplex checking.
+          duplexToTry.push_back(DuplexCandidate(k, j, iClass));
+          DEBUG(dbgs() << "adding pair:" << k << "," << j << ":"
+                       << MCB.getOperand(j).getInst()->getOpcode() << ","
+                       << MCB.getOperand(k).getInst()->getOpcode() << "\n");
+        } else {
+          DEBUG(dbgs() << "skipping pair: " << k << "," << j << ":"
+                       << MCB.getOperand(j).getInst()->getOpcode() << ","
+                       << MCB.getOperand(k).getInst()->getOpcode() << "\n");
+        }
+      }
+    }
+  }
+  return duplexToTry;
+}
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.cpp b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.cpp
index 7b0760e8318..2731278f0e4 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.cpp
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.cpp
@@ -16,6 +16,8 @@
 #include "Hexagon.h"
 #include "HexagonBaseInfo.h"
 
+#include "llvm/MC/MCContext.h"
+#include "llvm/MC/MCInstrInfo.h"
 #include "llvm/MC/MCSubtargetInfo.h"
 
 namespace llvm {
@@ -33,6 +35,32 @@ size_t HexagonMCInstrInfo::bundleSize(MCInst const &MCI) {
     return (1);
 }
 
+MCInst *HexagonMCInstrInfo::deriveDuplex(MCContext &Context, unsigned iClass,
+                                         MCInst const &inst0,
+                                         MCInst const &inst1) {
+  assert((iClass <= 0xf) && "iClass must have range of 0 to 0xf");
+  MCInst *duplexInst = new (Context) MCInst;
+  duplexInst->setOpcode(Hexagon::DuplexIClass0 + iClass);
+
+  MCInst *SubInst0 = new (Context) MCInst(deriveSubInst(inst0));
+  MCInst *SubInst1 = new (Context) MCInst(deriveSubInst(inst1));
+  duplexInst->addOperand(MCOperand::createInst(SubInst0));
+  duplexInst->addOperand(MCOperand::createInst(SubInst1));
+  return duplexInst;
+}
+
+MCInst const *HexagonMCInstrInfo::extenderForIndex(MCInst const &MCB,
+                                                   size_t Index) {
+  assert(Index <= bundleSize(MCB));
+  if (Index == 0)
+    return nullptr;
+  MCInst const *Inst =
+      MCB.getOperand(Index + bundleInstructionsOffset - 1).getInst();
+  if (isImmext(*Inst))
+    return Inst;
+  return nullptr;
+}
+
 HexagonII::MemAccessSize
 HexagonMCInstrInfo::getAccessSize(MCInstrInfo const &MCII, MCInst const &MCI) {
   const uint64_t F = HexagonMCInstrInfo::getDesc(MCII, MCI).TSFlags;
@@ -188,6 +216,10 @@ bool HexagonMCInstrInfo::hasImmExt(MCInst const &MCI) {
   return false;
 }
 
+bool HexagonMCInstrInfo::hasExtenderForIndex(MCInst const &MCB, size_t Index) {
+  return extenderForIndex(MCB, Index) != nullptr;
+}
+
 // Return whether the instruction is a legal new-value producer.
 bool HexagonMCInstrInfo::hasNewValue(MCInstrInfo const &MCII,
                                      MCInst const &MCI) {
@@ -214,6 +246,15 @@ bool HexagonMCInstrInfo::isCanon(MCInstrInfo const &MCII, MCInst const &MCI) {
           HexagonMCInstrInfo::getType(MCII, MCI) != HexagonII::TypeENDLOOP);
 }
 
+bool HexagonMCInstrInfo::isDblRegForSubInst(unsigned Reg) {
+  return ((Reg >= Hexagon::D0 && Reg <= Hexagon::D3) ||
+          (Reg >= Hexagon::D8 && Reg <= Hexagon::D11));
+}
+
+bool HexagonMCInstrInfo::isDuplex(MCInstrInfo const &MCII, MCInst const &MCI) {
+  return HexagonII::TypeDUPLEX == HexagonMCInstrInfo::getType(MCII, MCI);
+}
+
 // Return whether the instruction needs to be constant extended.
 // 1) Always return true if the instruction has 'isExtended' flag set.
 //
@@ -281,6 +322,15 @@ bool HexagonMCInstrInfo::isInnerLoop(MCInst const &MCI) {
   return (Flags & innerLoopMask) != 0;
 }
 
+bool HexagonMCInstrInfo::isIntReg(unsigned Reg) {
+  return (Reg >= Hexagon::R0 && Reg <= Hexagon::R31);
+}
+
+bool HexagonMCInstrInfo::isIntRegForSubInst(unsigned Reg) {
+  return ((Reg >= Hexagon::R0 && Reg <= Hexagon::R7) ||
+          (Reg >= Hexagon::R16 && Reg <= Hexagon::R23));
+}
+
 // Return whether the insn is a new-value consumer.
 bool HexagonMCInstrInfo::isNewValue(MCInstrInfo const &MCII,
                                     MCInst const &MCI) {
@@ -316,6 +366,10 @@ bool HexagonMCInstrInfo::isPredicatedTrue(MCInstrInfo const &MCII,
       !((F >> HexagonII::PredicatedFalsePos) & HexagonII::PredicatedFalseMask));
 }
 
+bool HexagonMCInstrInfo::isPredReg(unsigned Reg) {
+  return (Reg >= Hexagon::P0 && Reg <= Hexagon::P3_0);
+}
+
 bool HexagonMCInstrInfo::isPrefix(MCInstrInfo const &MCII, MCInst const &MCI) {
   return (HexagonMCInstrInfo::getType(MCII, MCI) == HexagonII::TypePREFIX);
 }
@@ -367,6 +421,20 @@ bool HexagonMCInstrInfo::prefersSlot3(MCInstrInfo const &MCII,
   return false;
 }
 
+void HexagonMCInstrInfo::replaceDuplex(MCContext &Context, MCInst &MCB,
+                                       DuplexCandidate Candidate) {
+  assert(Candidate.packetIndexI < MCB.size());
+  assert(Candidate.packetIndexJ < MCB.size());
+  assert(isBundle(MCB));
+  MCInst *Duplex =
+      deriveDuplex(Context, Candidate.iClass,
+                   *MCB.getOperand(Candidate.packetIndexJ).getInst(),
+                   *MCB.getOperand(Candidate.packetIndexI).getInst());
+  assert(Duplex != nullptr);
+  MCB.getOperand(Candidate.packetIndexI).setInst(Duplex);
+  MCB.erase(MCB.begin() + Candidate.packetIndexJ);
+}
+
 void HexagonMCInstrInfo::setInnerLoop(MCInst &MCI) {
   assert(isBundle(MCI));
   MCOperand &Operand = MCI.getOperand(0);
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.h b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.h
index 99970e2577c..f09dfd8cf47 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.h
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.h
@@ -15,15 +15,14 @@
 #define LLVM_LIB_TARGET_HEXAGON_MCTARGETDESC_HEXAGONMCINSTRINFO_H
 
 #include "llvm/MC/MCInst.h"
-#include "llvm/MC/MCInstrInfo.h"
-
-#include <bitset>
 
 namespace llvm {
+class MCContext;
 class MCInstrDesc;
 class MCInstrInfo;
 class MCInst;
 class MCOperand;
+class MCSubtargetInfo;
 namespace HexagonII {
 enum class MemAccessSize;
 }
@@ -48,6 +47,16 @@ size_t bundleSize(MCInst const &MCI);
 // Returns a iterator range of instructions in this bundle
 iterator_range<MCInst::const_iterator> bundleInstructions(MCInst const &MCI);
 
+// Return the extender for instruction at Index or nullptr if none
+MCInst const *extenderForIndex(MCInst const &MCB, size_t Index);
+
+// Create a duplex instruction given the two subinsts
+MCInst *deriveDuplex(MCContext &Context, unsigned iClass, MCInst const &inst0,
+                     MCInst const &inst1);
+
+// Convert this instruction in to a duplex subinst
+MCInst deriveSubInst(MCInst const &Inst);
+
 // Return memory access size
 HexagonII::MemAccessSize getAccessSize(MCInstrInfo const &MCII,
                                        MCInst const &MCI);
@@ -60,6 +69,13 @@ unsigned short getCExtOpNum(MCInstrInfo const &MCII, MCInst const &MCI);
 
 MCInstrDesc const &getDesc(MCInstrInfo const &MCII, MCInst const &MCI);
 
+// Return which duplex group this instruction belongs to
+unsigned getDuplexCandidateGroup(MCInst const &MI);
+
+// Return a list of all possible instruction duplex combinations
+SmallVector<DuplexCandidate, 8> getDuplexPossibilties(MCInstrInfo const &MCII,
+                                                      MCInst const &MCB);
+
 // Return the index of the extendable operand
 unsigned short getExtendableOp(MCInstrInfo const &MCII, MCInst const &MCI);
 
@@ -99,6 +115,9 @@ unsigned getType(MCInstrInfo const &MCII, MCInst const &MCI);
 unsigned getUnits(MCInstrInfo const &MCII, MCSubtargetInfo const &STI,
                   MCInst const &MCI);
 
+// Does the packet have an extender for the instruction at Index
+bool hasExtenderForIndex(MCInst const &MCB, size_t Index);
+
 bool hasImmExt(MCInst const &MCI);
 
 // Return whether the instruction is a legal new-value producer.
@@ -113,9 +132,24 @@ bool isBundle(MCInst const &MCI);
 // Return whether the insn is an actual insn.
 bool isCanon(MCInstrInfo const &MCII, MCInst const &MCI);
 
+// Return the duplex iclass given the two duplex classes
+unsigned iClassOfDuplexPair(unsigned Ga, unsigned Gb);
+
 // Return whether the instruction needs to be constant extended.
 bool isConstExtended(MCInstrInfo const &MCII, MCInst const &MCI);
 
+// Is this double register suitable for use in a duplex subinst
+bool isDblRegForSubInst(unsigned Reg);
+
+// Is this a duplex instruction
+bool isDuplex(MCInstrInfo const &MCII, MCInst const &MCI);
+
+// Can these instructions be duplexed
+bool isDuplexPair(MCInst const &MIa, MCInst const &MIb);
+
+// Can these duplex classes be combine in to a duplex instruction
+bool isDuplexPairMatch(unsigned Ga, unsigned Gb);
+
 // Return true if the insn may be extended based on the operand value.
 bool isExtendable(MCInstrInfo const &MCII, MCInst const &MCI);
 
@@ -131,6 +165,12 @@ bool isImmext(MCInst const &MCI);
 // Returns whether this bundle is an endloop0
 bool isInnerLoop(MCInst const &MCI);
 
+// Is this an integer register
+bool isIntReg(unsigned Reg);
+
+// Is this register suitable for use in a duplex subinst
+bool isIntRegForSubInst(unsigned Reg);
+
 // Return whether the insn is a new-value consumer.
 bool isNewValue(MCInstrInfo const &MCII, MCInst const &MCI);
 
@@ -138,6 +178,11 @@ bool isNewValue(MCInstrInfo const &MCII, MCInst const &MCI);
 bool isOperandExtended(MCInstrInfo const &MCII, MCInst const &MCI,
                        unsigned short OperandNum);
 
+// Can these two instructions be duplexed
+bool isOrderedDuplexPair(MCInstrInfo const &MCII, MCInst const &MIa,
+                         bool ExtendedA, MCInst const &MIb, bool ExtendedB,
+                         bool bisReversable);
+
 // Returns whether this bundle is an endloop1
 bool isOuterLoop(MCInst const &MCI);
 
@@ -147,6 +192,9 @@ bool isPredicated(MCInstrInfo const &MCII, MCInst const &MCI);
 // Return whether the predicate sense is true
 bool isPredicatedTrue(MCInstrInfo const &MCII, MCInst const &MCI);
 
+// Is this a predicate register
+bool isPredReg(unsigned Reg);
+
 // Return whether the insn is a prefix.
 bool isPrefix(MCInstrInfo const &MCII, MCInst const &MCI);
 
@@ -164,11 +212,17 @@ void padEndloop(MCInst &MCI);
 
 bool prefersSlot3(MCInstrInfo const &MCII, MCInst const &MCI);
 
+// Replace the instructions inside MCB, represented by Candidate
+void replaceDuplex(MCContext &Context, MCInst &MCB, DuplexCandidate Candidate);
+
 // Marks a bundle as endloop0
 void setInnerLoop(MCInst &MCI);
 
 // Marks a bundle as endloop1
 void setOuterLoop(MCInst &MCI);
+
+// Would duplexing this instruction create a requirement to extend
+bool subInstWouldBeExtended(MCInst const &potentialDuplex);
 }
 }
 
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCShuffler.cpp b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCShuffler.cpp
index 37cf3754870..8e70280c1a0 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCShuffler.cpp
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCShuffler.cpp
@@ -147,6 +147,61 @@ bool llvm::HexagonMCShuffle(MCInstrInfo const &MCII, MCSubtargetInfo const &STI,
   return true;
 }
 
+unsigned
+llvm::HexagonMCShuffle(MCInstrInfo const &MCII, MCSubtargetInfo const &STI,
+                       MCContext &Context, MCInst &MCB,
+                       SmallVector<DuplexCandidate, 8> possibleDuplexes) {
+
+  if (DisableShuffle)
+    return HexagonShuffler::SHUFFLE_SUCCESS;
+
+  if (!HexagonMCInstrInfo::bundleSize(MCB)) {
+    // There once was a bundle:
+    //    BUNDLE %D2<imp-def>, %R4<imp-def>, %R5<imp-def>, %D7<imp-def>, ...
+    //      * %D2<def> = IMPLICIT_DEF; flags:
+    //      * %D7<def> = IMPLICIT_DEF; flags:
+    // After the IMPLICIT_DEFs were removed by the asm printer, the bundle
+    // became empty.
+    DEBUG(dbgs() << "Skipping empty bundle");
+    return HexagonShuffler::SHUFFLE_SUCCESS;
+  } else if (!HexagonMCInstrInfo::isBundle(MCB)) {
+    DEBUG(dbgs() << "Skipping stand-alone insn");
+    return HexagonShuffler::SHUFFLE_SUCCESS;
+  }
+
+  bool doneShuffling = false;
+  unsigned shuffleError;
+  while (possibleDuplexes.size() > 0 && (!doneShuffling)) {
+    // case of Duplex Found
+    DuplexCandidate duplexToTry = possibleDuplexes.pop_back_val();
+    MCInst Attempt(MCB);
+    HexagonMCInstrInfo::replaceDuplex(Context, Attempt, duplexToTry);
+    HexagonMCShuffler MCS(MCII, STI, Attempt); // copy packet to the shuffler
+    if (MCS.size() == 1) {                     // case of one duplex
+      // copy the created duplex in the shuffler to the bundle
+      MCS.copyTo(MCB);
+      doneShuffling = true;
+      return HexagonShuffler::SHUFFLE_SUCCESS;
+    }
+    // try shuffle with this duplex
+    doneShuffling = MCS.reshuffleTo(MCB);
+    shuffleError = MCS.getError();
+
+    if (doneShuffling)
+      break;
+  }
+
+  if (doneShuffling == false) {
+    HexagonMCShuffler MCS(MCII, STI, MCB);
+    doneShuffling = MCS.reshuffleTo(MCB); // shuffle
+    shuffleError = MCS.getError();
+  }
+  if (!doneShuffling)
+    return shuffleError;
+
+  return HexagonShuffler::SHUFFLE_SUCCESS;
+}
+
 bool llvm::HexagonMCShuffle(MCInstrInfo const &MCII, MCSubtargetInfo const &STI,
                             MCInst &MCB, MCInst const *AddMI, int fixupCount) {
   if (!HexagonMCInstrInfo::isBundle(MCB) || !AddMI)