[llvm] r241683 - [Hexagon] Generate "insert" instructions more aggressively

Krzysztof Parzyszek kparzysz at codeaurora.org
Wed Jul 8 07:47:35 PDT 2015


Author: kparzysz
Date: Wed Jul  8 09:47:34 2015
New Revision: 241683

URL: http://llvm.org/viewvc/llvm-project?rev=241683&view=rev
Log:
[Hexagon] Generate "insert" instructions more aggressively


Added:
    llvm/trunk/lib/Target/Hexagon/HexagonGenInsert.cpp
    llvm/trunk/test/CodeGen/Hexagon/insert-basic.ll
Modified:
    llvm/trunk/lib/Target/Hexagon/CMakeLists.txt
    llvm/trunk/lib/Target/Hexagon/HexagonTargetMachine.cpp

Modified: llvm/trunk/lib/Target/Hexagon/CMakeLists.txt
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/Hexagon/CMakeLists.txt?rev=241683&r1=241682&r2=241683&view=diff
==============================================================================
--- llvm/trunk/lib/Target/Hexagon/CMakeLists.txt (original)
+++ llvm/trunk/lib/Target/Hexagon/CMakeLists.txt Wed Jul  8 09:47:34 2015
@@ -21,6 +21,7 @@ add_llvm_target(HexagonCodeGen
   HexagonExpandPredSpillCode.cpp
   HexagonFixupHwLoops.cpp
   HexagonFrameLowering.cpp
+  HexagonGenInsert.cpp
   HexagonHardwareLoops.cpp
   HexagonInstrInfo.cpp
   HexagonISelDAGToDAG.cpp

Added: llvm/trunk/lib/Target/Hexagon/HexagonGenInsert.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/Hexagon/HexagonGenInsert.cpp?rev=241683&view=auto
==============================================================================
--- llvm/trunk/lib/Target/Hexagon/HexagonGenInsert.cpp (added)
+++ llvm/trunk/lib/Target/Hexagon/HexagonGenInsert.cpp Wed Jul  8 09:47:34 2015
@@ -0,0 +1,1598 @@
+//===--- HexagonGenInsert.cpp ---------------------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "hexinsert"
+
+#include "llvm/Pass.h"
+#include "llvm/PassRegistry.h"
+#include "llvm/ADT/BitVector.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/DenseSet.h"
+#include "llvm/ADT/PostOrderIterator.h"
+#include "llvm/CodeGen/MachineDominators.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Support/Timer.h"
+#include "llvm/Target/TargetMachine.h"
+#include "llvm/Target/TargetRegisterInfo.h"
+
+#include "Hexagon.h"
+#include "HexagonRegisterInfo.h"
+#include "HexagonTargetMachine.h"
+#include "HexagonBitTracker.h"
+
+#include <map>
+#include <vector>
+
+using namespace llvm;
+
+static cl::opt<unsigned> VRegIndexCutoff("insert-vreg-cutoff", cl::init(~0U),
+  cl::Hidden, cl::ZeroOrMore, cl::desc("Vreg# cutoff for insert generation."));
+// The distance cutoff is selected based on the precheckin-perf results:
+// cutoffs 20, 25, 35, and 40 are worse than 30.
+static cl::opt<unsigned> VRegDistCutoff("insert-dist-cutoff", cl::init(30U),
+  cl::Hidden, cl::ZeroOrMore, cl::desc("Vreg distance cutoff for insert "
+  "generation."));
+
+static cl::opt<bool> OptTiming("insert-timing", cl::init(false), cl::Hidden,
+  cl::ZeroOrMore, cl::desc("Enable timing of insert generation"));
+static cl::opt<bool> OptTimingDetail("insert-timing-detail", cl::init(false),
+  cl::Hidden, cl::ZeroOrMore, cl::desc("Enable detailed timing of insert "
+  "generation"));
+
+static cl::opt<bool> OptSelectAll0("insert-all0", cl::init(false), cl::Hidden,
+  cl::ZeroOrMore);
+static cl::opt<bool> OptSelectHas0("insert-has0", cl::init(false), cl::Hidden,
+  cl::ZeroOrMore);
+// Whether to construct constant values via "insert". Could eliminate constant
+// extenders, but often not practical.
+static cl::opt<bool> OptConst("insert-const", cl::init(false), cl::Hidden,
+  cl::ZeroOrMore);
+
+namespace {
+  // The preprocessor gets confused when the DEBUG macro is passed larger
+  // chunks of code. Use this function to detect debugging.
+  inline bool isDebug() {
+#ifndef NDEBUG
+    return ::llvm::DebugFlag && ::llvm::isCurrentDebugType(DEBUG_TYPE);
+#else
+    return false;
+#endif
+  }
+}
+
+
+namespace {
+  // Set of virtual registers, based on BitVector.
+  struct RegisterSet : private BitVector {
+    RegisterSet() : BitVector() {}
+    explicit RegisterSet(unsigned s, bool t = false) : BitVector(s, t) {}
+    RegisterSet(const RegisterSet &RS) : BitVector(RS) {}
+
+    using BitVector::clear;
+
+    unsigned find_first() const {
+      int First = BitVector::find_first();
+      if (First < 0)
+        return 0;
+      return x2v(First);
+    }
+
+    unsigned find_next(unsigned Prev) const {
+      int Next = BitVector::find_next(v2x(Prev));
+      if (Next < 0)
+        return 0;
+      return x2v(Next);
+    }
+
+    RegisterSet &insert(unsigned R) {
+      unsigned Idx = v2x(R);
+      ensure(Idx);
+      return static_cast<RegisterSet&>(BitVector::set(Idx));
+    }
+    RegisterSet &remove(unsigned R) {
+      unsigned Idx = v2x(R);
+      if (Idx >= size())
+        return *this;
+      return static_cast<RegisterSet&>(BitVector::reset(Idx));
+    }
+
+    RegisterSet &insert(const RegisterSet &Rs) {
+      return static_cast<RegisterSet&>(BitVector::operator|=(Rs));
+    }
+    RegisterSet &remove(const RegisterSet &Rs) {
+      return static_cast<RegisterSet&>(BitVector::reset(Rs));
+    }
+
+    reference operator[](unsigned R) {
+      unsigned Idx = v2x(R);
+      ensure(Idx);
+      return BitVector::operator[](Idx);
+    }
+    bool operator[](unsigned R) const {
+      unsigned Idx = v2x(R);
+      assert(Idx < size());
+      return BitVector::operator[](Idx);
+    }
+    bool has(unsigned R) const {
+      unsigned Idx = v2x(R);
+      if (Idx >= size())
+        return false;
+      return BitVector::test(Idx);
+    }
+
+    bool empty() const {
+      return !BitVector::any();
+    }
+    bool includes(const RegisterSet &Rs) const {
+      // A.BitVector::test(B)  <=>  A-B != {}
+      return !Rs.BitVector::test(*this);
+    }
+    bool intersects(const RegisterSet &Rs) const {
+      return BitVector::anyCommon(Rs);
+    }
+
+  private:
+    void ensure(unsigned Idx) {
+      if (size() <= Idx)
+        resize(std::max(Idx+1, 32U));
+    }
+    static inline unsigned v2x(unsigned v) {
+      return TargetRegisterInfo::virtReg2Index(v);
+    }
+    static inline unsigned x2v(unsigned x) {
+      return TargetRegisterInfo::index2VirtReg(x);
+    }
+  };
+
+
+  struct PrintRegSet {
+    PrintRegSet(const RegisterSet &S, const TargetRegisterInfo *RI)
+      : RS(S), TRI(RI) {}
+    friend raw_ostream &operator<< (raw_ostream &OS,
+          const PrintRegSet &P);
+  private:
+    const RegisterSet &RS;
+    const TargetRegisterInfo *TRI;
+  };
+
+  raw_ostream &operator<< (raw_ostream &OS, const PrintRegSet &P) {
+    OS << '{';
+    for (unsigned R = P.RS.find_first(); R; R = P.RS.find_next(R))
+      OS << ' ' << PrintReg(R, P.TRI);
+    OS << " }";
+    return OS;
+  }
+}
+
+
+namespace {
+  // A convenience class to associate unsigned numbers (such as virtual
+  // registers) with unsigned numbers.
+  struct UnsignedMap : public DenseMap<unsigned,unsigned> {
+    UnsignedMap() : BaseType() {}
+  private:
+    typedef DenseMap<unsigned,unsigned> BaseType;
+  };
+
+  // A utility to establish an ordering between virtual registers:
+  // VRegA < VRegB  <=>  RegisterOrdering[VRegA] < RegisterOrdering[VRegB]
+  // This is meant as a cache for the ordering of virtual registers defined
+  // by a potentially expensive comparison function, or obtained by a proce-
+  // dure that should not be repeated each time two registers are compared.
+  struct RegisterOrdering : public UnsignedMap {
+    RegisterOrdering() : UnsignedMap() {}
+    unsigned operator[](unsigned VR) const {
+      const_iterator F = find(VR);
+      assert(F != end());
+      return F->second;
+    }
+    // Add operator(), so that objects of this class can be used as
+    // comparators in std::sort et al.
+    bool operator() (unsigned VR1, unsigned VR2) const {
+      return operator[](VR1) < operator[](VR2);
+    }
+  };
+}
+
+
+namespace {
+  // Ordering of bit values. This class does not have operator[], but
+  // is supplies a comparison operator() for use in std:: algorithms.
+  // The order is as follows:
+  // - 0 < 1 < ref
+  // - ref1 < ref2, if ord(ref1.Reg) < ord(ref2.Reg),
+  //   or ord(ref1.Reg) == ord(ref2.Reg), and ref1.Pos < ref2.Pos.
+  struct BitValueOrdering {
+    BitValueOrdering(const RegisterOrdering &RB) : BaseOrd(RB) {}
+    bool operator() (const BitTracker::BitValue &V1,
+          const BitTracker::BitValue &V2) const;
+    const RegisterOrdering &BaseOrd;
+  };
+}
+
+
+bool BitValueOrdering::operator() (const BitTracker::BitValue &V1,
+      const BitTracker::BitValue &V2) const {
+  if (V1 == V2)
+    return false;
+  // V1==0 => true, V2==0 => false
+  if (V1.is(0) || V2.is(0))
+    return V1.is(0);
+  // Neither of V1,V2 is 0, and V1!=V2.
+  // V2==1 => false, V1==1 => true
+  if (V2.is(1) || V1.is(1))
+    return !V2.is(1);
+  // Both V1,V2 are refs.
+  unsigned Ind1 = BaseOrd[V1.RefI.Reg], Ind2 = BaseOrd[V2.RefI.Reg];
+  if (Ind1 != Ind2)
+    return Ind1 < Ind2;
+  // If V1.Pos==V2.Pos
+  assert(V1.RefI.Pos != V2.RefI.Pos && "Bit values should be different");
+  return V1.RefI.Pos < V2.RefI.Pos;
+}
+
+
+namespace {
+  // Cache for the BitTracker's cell map. Map lookup has a logarithmic
+  // complexity, this class will memoize the lookup results to reduce
+  // the access time for repeated lookups of the same cell.
+  struct CellMapShadow {
+    CellMapShadow(const BitTracker &T) : BT(T) {}
+    const BitTracker::RegisterCell &lookup(unsigned VR) {
+      unsigned RInd = TargetRegisterInfo::virtReg2Index(VR);
+      // Grow the vector to at least 32 elements.
+      if (RInd >= CVect.size())
+        CVect.resize(std::max(RInd+16, 32U), 0);
+      const BitTracker::RegisterCell *CP = CVect[RInd];
+      if (CP == 0)
+        CP = CVect[RInd] = &BT.lookup(VR);
+      return *CP;
+    }
+
+    const BitTracker &BT;
+
+  private:
+    typedef std::vector<const BitTracker::RegisterCell*> CellVectType;
+    CellVectType CVect;
+  };
+}
+
+
+namespace {
+  // Comparator class for lexicographic ordering of virtual registers
+  // according to the corresponding BitTracker::RegisterCell objects.
+  struct RegisterCellLexCompare {
+    RegisterCellLexCompare(const BitValueOrdering &BO, CellMapShadow &M)
+      : BitOrd(BO), CM(M) {}
+    bool operator() (unsigned VR1, unsigned VR2) const;
+  private:
+    const BitValueOrdering &BitOrd;
+    CellMapShadow &CM;
+  };
+
+  // Comparator class for lexicographic ordering of virtual registers
+  // according to the specified bits of the corresponding BitTracker::
+  // RegisterCell objects.
+  // Specifically, this class will be used to compare bit B of a register
+  // cell for a selected virtual register R with bit N of any register
+  // other than R.
+  struct RegisterCellBitCompareSel {
+    RegisterCellBitCompareSel(unsigned R, unsigned B, unsigned N,
+          const BitValueOrdering &BO, CellMapShadow &M)
+      : SelR(R), SelB(B), BitN(N), BitOrd(BO), CM(M) {}
+    bool operator() (unsigned VR1, unsigned VR2) const;
+  private:
+    const unsigned SelR, SelB;
+    const unsigned BitN;
+    const BitValueOrdering &BitOrd;
+    CellMapShadow &CM;
+  };
+}
+
+
+bool RegisterCellLexCompare::operator() (unsigned VR1, unsigned VR2) const {
+  // Ordering of registers, made up from two given orderings:
+  // - the ordering of the register numbers, and
+  // - the ordering of register cells.
+  // Def. R1 < R2 if:
+  // - cell(R1) < cell(R2), or
+  // - cell(R1) == cell(R2), and index(R1) < index(R2).
+  //
+  // For register cells, the ordering is lexicographic, with index 0 being
+  // the most significant.
+  if (VR1 == VR2)
+    return false;
+
+  const BitTracker::RegisterCell &RC1 = CM.lookup(VR1), &RC2 = CM.lookup(VR2);
+  uint16_t W1 = RC1.width(), W2 = RC2.width();
+  for (uint16_t i = 0, w = std::min(W1, W2); i < w; ++i) {
+    const BitTracker::BitValue &V1 = RC1[i], &V2 = RC2[i];
+    if (V1 != V2)
+      return BitOrd(V1, V2);
+  }
+  // Cells are equal up until the common length.
+  if (W1 != W2)
+    return W1 < W2;
+
+  return BitOrd.BaseOrd[VR1] < BitOrd.BaseOrd[VR2];
+}
+
+
+bool RegisterCellBitCompareSel::operator() (unsigned VR1, unsigned VR2) const {
+  if (VR1 == VR2)
+    return false;
+  const BitTracker::RegisterCell &RC1 = CM.lookup(VR1);
+  const BitTracker::RegisterCell &RC2 = CM.lookup(VR2);
+  uint16_t W1 = RC1.width(), W2 = RC2.width();
+  uint16_t Bit1 = (VR1 == SelR) ? SelB : BitN;
+  uint16_t Bit2 = (VR2 == SelR) ? SelB : BitN;
+  // If Bit1 exceeds the width of VR1, then:
+  // - return false, if at the same time Bit2 exceeds VR2, or
+  // - return true, otherwise.
+  // (I.e. "a bit value that does not exist is less than any bit value
+  // that does exist".)
+  if (W1 <= Bit1)
+    return Bit2 < W2;
+  // If Bit1 is within VR1, but Bit2 is not within VR2, return false.
+  if (W2 <= Bit2)
+    return false;
+
+  const BitTracker::BitValue &V1 = RC1[Bit1], V2 = RC2[Bit2];
+  if (V1 != V2)
+    return BitOrd(V1, V2);
+  return false;
+}
+
+
+namespace {
+  class OrderedRegisterList {
+    typedef std::vector<unsigned> ListType;
+  public:
+    OrderedRegisterList(const RegisterOrdering &RO) : Ord(RO) {}
+    void insert(unsigned VR);
+    void remove(unsigned VR);
+    unsigned operator[](unsigned Idx) const {
+      assert(Idx < Seq.size());
+      return Seq[Idx];
+    }
+    unsigned size() const {
+      return Seq.size();
+    }
+
+    typedef ListType::iterator iterator;
+    typedef ListType::const_iterator const_iterator;
+    iterator begin() { return Seq.begin(); }
+    iterator end() { return Seq.end(); }
+    const_iterator begin() const { return Seq.begin(); }
+    const_iterator end() const { return Seq.end(); }
+
+    // Convenience function to convert an iterator to the corresponding index.
+    unsigned idx(iterator It) const { return It-begin(); }
+  private:
+    ListType Seq;
+    const RegisterOrdering &Ord;
+  };
+
+
+  struct PrintORL {
+    PrintORL(const OrderedRegisterList &L, const TargetRegisterInfo *RI)
+      : RL(L), TRI(RI) {}
+    friend raw_ostream &operator<< (raw_ostream &OS, const PrintORL &P);
+  private:
+    const OrderedRegisterList &RL;
+    const TargetRegisterInfo *TRI;
+  };
+
+  raw_ostream &operator<< (raw_ostream &OS, const PrintORL &P) {
+    OS << '(';
+    OrderedRegisterList::const_iterator B = P.RL.begin(), E = P.RL.end();
+    for (OrderedRegisterList::const_iterator I = B; I != E; ++I) {
+      if (I != B)
+        OS << ", ";
+      OS << PrintReg(*I, P.TRI);
+    }
+    OS << ')';
+    return OS;
+  }
+}
+
+
+void OrderedRegisterList::insert(unsigned VR) {
+  iterator L = std::lower_bound(Seq.begin(), Seq.end(), VR, Ord);
+  if (L == Seq.end())
+    Seq.push_back(VR);
+  else
+    Seq.insert(L, VR);
+}
+
+
+void OrderedRegisterList::remove(unsigned VR) {
+  iterator L = std::lower_bound(Seq.begin(), Seq.end(), VR, Ord);
+  assert(L != Seq.end());
+  Seq.erase(L);
+}
+
+
+namespace {
+  // A record of the insert form. The fields correspond to the operands
+  // of the "insert" instruction:
+  // ... = insert(SrcR, InsR, #Wdh, #Off)
+  struct IFRecord {
+    IFRecord(unsigned SR = 0, unsigned IR = 0, uint16_t W = 0, uint16_t O = 0)
+      : SrcR(SR), InsR(IR), Wdh(W), Off(O) {}
+    unsigned SrcR, InsR;
+    uint16_t Wdh, Off;
+  };
+
+  struct PrintIFR {
+    PrintIFR(const IFRecord &R, const TargetRegisterInfo *RI)
+      : IFR(R), TRI(RI) {}
+  private:
+    const IFRecord &IFR;
+    const TargetRegisterInfo *TRI;
+    friend raw_ostream &operator<< (raw_ostream &OS, const PrintIFR &P);
+  };
+
+  raw_ostream &operator<< (raw_ostream &OS, const PrintIFR &P) {
+    unsigned SrcR = P.IFR.SrcR, InsR = P.IFR.InsR;
+    OS << '(' << PrintReg(SrcR, P.TRI) << ',' << PrintReg(InsR, P.TRI)
+       << ",#" << P.IFR.Wdh << ",#" << P.IFR.Off << ')';
+    return OS;
+  }
+
+  typedef std::pair<IFRecord,RegisterSet> IFRecordWithRegSet;
+}
+
+
+namespace llvm {
+  void initializeHexagonGenInsertPass(PassRegistry&);
+  FunctionPass *createHexagonGenInsert();
+}
+
+
+namespace {
+  class HexagonGenInsert : public MachineFunctionPass {
+  public:
+    static char ID;
+    HexagonGenInsert() : MachineFunctionPass(ID), HII(0), HRI(0) {
+      initializeHexagonGenInsertPass(*PassRegistry::getPassRegistry());
+    }
+    virtual const char *getPassName() const {
+      return "Hexagon generate \"insert\" instructions";
+    }
+    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+      AU.addRequired<MachineDominatorTree>();
+      AU.addPreserved<MachineDominatorTree>();
+      MachineFunctionPass::getAnalysisUsage(AU);
+    }
+    virtual bool runOnMachineFunction(MachineFunction &MF);
+
+  private:
+    typedef DenseMap<std::pair<unsigned,unsigned>,unsigned> PairMapType;
+
+    void buildOrderingMF(RegisterOrdering &RO) const;
+    void buildOrderingBT(RegisterOrdering &RB, RegisterOrdering &RO) const;
+    bool isIntClass(const TargetRegisterClass *RC) const;
+    bool isConstant(unsigned VR) const;
+    bool isSmallConstant(unsigned VR) const;
+    bool isValidInsertForm(unsigned DstR, unsigned SrcR, unsigned InsR,
+          uint16_t L, uint16_t S) const;
+    bool findSelfReference(unsigned VR) const;
+    bool findNonSelfReference(unsigned VR) const;
+    void getInstrDefs(const MachineInstr *MI, RegisterSet &Defs) const;
+    void getInstrUses(const MachineInstr *MI, RegisterSet &Uses) const;
+    unsigned distance(const MachineBasicBlock *FromB,
+          const MachineBasicBlock *ToB, const UnsignedMap &RPO,
+          PairMapType &M) const;
+    unsigned distance(MachineBasicBlock::const_iterator FromI,
+          MachineBasicBlock::const_iterator ToI, const UnsignedMap &RPO,
+          PairMapType &M) const;
+    bool findRecordInsertForms(unsigned VR, OrderedRegisterList &AVs);
+    void collectInBlock(MachineBasicBlock *B, OrderedRegisterList &AVs);
+    void findRemovableRegisters(unsigned VR, IFRecord IF,
+          RegisterSet &RMs) const;
+    void computeRemovableRegisters();
+
+    void pruneEmptyLists();
+    void pruneCoveredSets(unsigned VR);
+    void pruneUsesTooFar(unsigned VR, const UnsignedMap &RPO, PairMapType &M);
+    void pruneRegCopies(unsigned VR);
+    void pruneCandidates();
+    void selectCandidates();
+    bool generateInserts();
+
+    bool removeDeadCode(MachineDomTreeNode *N);
+
+    // IFRecord coupled with a set of potentially removable registers:
+    typedef std::vector<IFRecordWithRegSet> IFListType;
+    typedef DenseMap<unsigned,IFListType> IFMapType;  // vreg -> IFListType
+
+    void dump_map() const;
+
+    const HexagonInstrInfo *HII;
+    const HexagonRegisterInfo *HRI;
+
+    MachineFunction *MFN;
+    MachineRegisterInfo *MRI;
+    MachineDominatorTree *MDT;
+    CellMapShadow *CMS;
+
+    RegisterOrdering BaseOrd;
+    RegisterOrdering CellOrd;
+    IFMapType IFMap;
+  };
+
+  char HexagonGenInsert::ID = 0;
+}
+
+
+void HexagonGenInsert::dump_map() const {
+  typedef IFMapType::const_iterator iterator;
+  for (iterator I = IFMap.begin(), E = IFMap.end(); I != E; ++I) {
+    dbgs() << "  " << PrintReg(I->first, HRI) << ":\n";
+    const IFListType &LL = I->second;
+    for (unsigned i = 0, n = LL.size(); i < n; ++i)
+      dbgs() << "    " << PrintIFR(LL[i].first, HRI) << ", "
+             << PrintRegSet(LL[i].second, HRI) << '\n';
+  }
+}
+
+
+void HexagonGenInsert::buildOrderingMF(RegisterOrdering &RO) const {
+  unsigned Index = 0;
+  typedef MachineFunction::const_iterator mf_iterator;
+  for (mf_iterator A = MFN->begin(), Z = MFN->end(); A != Z; ++A) {
+    const MachineBasicBlock &B = *A;
+    if (!CMS->BT.reached(&B))
+      continue;
+    typedef MachineBasicBlock::const_iterator mb_iterator;
+    for (mb_iterator I = B.begin(), E = B.end(); I != E; ++I) {
+      const MachineInstr *MI = &*I;
+      for (unsigned i = 0, n = MI->getNumOperands(); i < n; ++i) {
+        const MachineOperand &MO = MI->getOperand(i);
+        if (MO.isReg() && MO.isDef()) {
+          unsigned R = MO.getReg();
+          assert(MO.getSubReg() == 0 && "Unexpected subregister in definition");
+          if (TargetRegisterInfo::isVirtualRegister(R))
+            RO.insert(std::make_pair(R, Index++));
+        }
+      }
+    }
+  }
+  // Since some virtual registers may have had their def and uses eliminated,
+  // they are no longer referenced in the code, and so they will not appear
+  // in the map.
+}
+
+
+void HexagonGenInsert::buildOrderingBT(RegisterOrdering &RB,
+      RegisterOrdering &RO) const {
+  // Create a vector of all virtual registers (collect them from the base
+  // ordering RB), and then sort it using the RegisterCell comparator.
+  BitValueOrdering BVO(RB);
+  RegisterCellLexCompare LexCmp(BVO, *CMS);
+  typedef std::vector<unsigned> SortableVectorType;
+  SortableVectorType VRs;
+  for (RegisterOrdering::iterator I = RB.begin(), E = RB.end(); I != E; ++I)
+    VRs.push_back(I->first);
+  std::sort(VRs.begin(), VRs.end(), LexCmp);
+  // Transfer the results to the outgoing register ordering.
+  for (unsigned i = 0, n = VRs.size(); i < n; ++i)
+    RO.insert(std::make_pair(VRs[i], i));
+}
+
+
+inline bool HexagonGenInsert::isIntClass(const TargetRegisterClass *RC) const {
+  return RC == &Hexagon::IntRegsRegClass || RC == &Hexagon::DoubleRegsRegClass;
+}
+
+
+bool HexagonGenInsert::isConstant(unsigned VR) const {
+  const BitTracker::RegisterCell &RC = CMS->lookup(VR);
+  uint16_t W = RC.width();
+  for (uint16_t i = 0; i < W; ++i) {
+    const BitTracker::BitValue &BV = RC[i];
+    if (BV.is(0) || BV.is(1))
+      continue;
+    return false;
+  }
+  return true;
+}
+
+
+bool HexagonGenInsert::isSmallConstant(unsigned VR) const {
+  const BitTracker::RegisterCell &RC = CMS->lookup(VR);
+  uint16_t W = RC.width();
+  if (W > 64)
+    return false;
+  uint64_t V = 0, B = 1;
+  for (uint16_t i = 0; i < W; ++i) {
+    const BitTracker::BitValue &BV = RC[i];
+    if (BV.is(1))
+      V |= B;
+    else if (!BV.is(0))
+      return false;
+    B <<= 1;
+  }
+
+  // For 32-bit registers, consider: Rd = #s16.
+  if (W == 32)
+    return isInt<16>(V);
+
+  // For 64-bit registers, it's Rdd = #s8 or Rdd = combine(#s8,#s8)
+  return isInt<8>(Lo_32(V)) && isInt<8>(Hi_32(V));
+}
+
+
+bool HexagonGenInsert::isValidInsertForm(unsigned DstR, unsigned SrcR,
+      unsigned InsR, uint16_t L, uint16_t S) const {
+  const TargetRegisterClass *DstRC = MRI->getRegClass(DstR);
+  const TargetRegisterClass *SrcRC = MRI->getRegClass(SrcR);
+  const TargetRegisterClass *InsRC = MRI->getRegClass(InsR);
+  // Only integet (32-/64-bit) register classes.
+  if (!isIntClass(DstRC) || !isIntClass(SrcRC) || !isIntClass(InsRC))
+    return false;
+  // The "source" register must be of the same class as DstR.
+  if (DstRC != SrcRC)
+    return false;
+  if (DstRC == InsRC)
+    return true;
+  // A 64-bit register can only be generated from other 64-bit registers.
+  if (DstRC == &Hexagon::DoubleRegsRegClass)
+    return false;
+  // Otherwise, the L and S cannot span 32-bit word boundary.
+  if (S < 32 && S+L > 32)
+    return false;
+  return true;
+}
+
+
+bool HexagonGenInsert::findSelfReference(unsigned VR) const {
+  const BitTracker::RegisterCell &RC = CMS->lookup(VR);
+  for (uint16_t i = 0, w = RC.width(); i < w; ++i) {
+    const BitTracker::BitValue &V = RC[i];
+    if (V.Type == BitTracker::BitValue::Ref && V.RefI.Reg == VR)
+      return true;
+  }
+  return false;
+}
+
+
+bool HexagonGenInsert::findNonSelfReference(unsigned VR) const {
+  BitTracker::RegisterCell RC = CMS->lookup(VR);
+  for (uint16_t i = 0, w = RC.width(); i < w; ++i) {
+    const BitTracker::BitValue &V = RC[i];
+    if (V.Type == BitTracker::BitValue::Ref && V.RefI.Reg != VR)
+      return true;
+  }
+  return false;
+}
+
+
+void HexagonGenInsert::getInstrDefs(const MachineInstr *MI,
+      RegisterSet &Defs) const {
+  for (unsigned i = 0, n = MI->getNumOperands(); i < n; ++i) {
+    const MachineOperand &MO = MI->getOperand(i);
+    if (!MO.isReg() || !MO.isDef())
+      continue;
+    unsigned R = MO.getReg();
+    if (!TargetRegisterInfo::isVirtualRegister(R))
+      continue;
+    Defs.insert(R);
+  }
+}
+
+
+void HexagonGenInsert::getInstrUses(const MachineInstr *MI,
+      RegisterSet &Uses) const {
+  for (unsigned i = 0, n = MI->getNumOperands(); i < n; ++i) {
+    const MachineOperand &MO = MI->getOperand(i);
+    if (!MO.isReg() || !MO.isUse())
+      continue;
+    unsigned R = MO.getReg();
+    if (!TargetRegisterInfo::isVirtualRegister(R))
+      continue;
+    Uses.insert(R);
+  }
+}
+
+
+unsigned HexagonGenInsert::distance(const MachineBasicBlock *FromB,
+      const MachineBasicBlock *ToB, const UnsignedMap &RPO,
+      PairMapType &M) const {
+  // Forward distance from the end of a block to the beginning of it does
+  // not make sense. This function should not be called with FromB == ToB.
+  assert(FromB != ToB);
+
+  unsigned FromN = FromB->getNumber(), ToN = ToB->getNumber();
+  // If we have already computed it, return the cached result.
+  PairMapType::iterator F = M.find(std::make_pair(FromN, ToN));
+  if (F != M.end())
+    return F->second;
+  unsigned ToRPO = RPO.lookup(ToN);
+
+  unsigned MaxD = 0;
+  typedef MachineBasicBlock::const_pred_iterator pred_iterator;
+  for (pred_iterator I = ToB->pred_begin(), E = ToB->pred_end(); I != E; ++I) {
+    const MachineBasicBlock *PB = *I;
+    // Skip back edges. Also, if FromB is a predecessor of ToB, the distance
+    // along that path will be 0, and we don't need to do any calculations
+    // on it.
+    if (PB == FromB || RPO.lookup(PB->getNumber()) >= ToRPO)
+      continue;
+    unsigned D = PB->size() + distance(FromB, PB, RPO, M);
+    if (D > MaxD)
+      MaxD = D;
+  }
+
+  // Memoize the result for later lookup.
+  M.insert(std::make_pair(std::make_pair(FromN, ToN), MaxD));
+  return MaxD;
+}
+
+
+unsigned HexagonGenInsert::distance(MachineBasicBlock::const_iterator FromI,
+      MachineBasicBlock::const_iterator ToI, const UnsignedMap &RPO,
+      PairMapType &M) const {
+  const MachineBasicBlock *FB = FromI->getParent(), *TB = ToI->getParent();
+  if (FB == TB)
+    return std::distance(FromI, ToI);
+  unsigned D1 = std::distance(TB->begin(), ToI);
+  unsigned D2 = distance(FB, TB, RPO, M);
+  unsigned D3 = std::distance(FromI, FB->end());
+  return D1+D2+D3;
+}
+
+
+bool HexagonGenInsert::findRecordInsertForms(unsigned VR,
+      OrderedRegisterList &AVs) {
+  if (isDebug()) {
+    dbgs() << LLVM_FUNCTION_NAME << ": " << PrintReg(VR, HRI)
+           << "  AVs: " << PrintORL(AVs, HRI) << "\n";
+  }
+  if (AVs.size() == 0)
+    return false;
+
+  typedef OrderedRegisterList::iterator iterator;
+  BitValueOrdering BVO(BaseOrd);
+  const BitTracker::RegisterCell &RC = CMS->lookup(VR);
+  uint16_t W = RC.width();
+
+  typedef std::pair<unsigned,uint16_t> RSRecord;  // (reg,shift)
+  typedef std::vector<RSRecord> RSListType;
+  // Have a map, with key being the matching prefix length, and the value
+  // being the list of pairs (R,S), where R's prefix matches VR at S.
+  // (DenseMap<uint16_t,RSListType> fails to instantiate.)
+  typedef DenseMap<unsigned,RSListType> LRSMapType;
+  LRSMapType LM;
+
+  // Conceptually, rotate the cell RC right (i.e. towards the LSB) by S,
+  // and find matching prefixes from AVs with the rotated RC. Such a prefix
+  // would match a string of bits (of length L) in RC starting at S.
+  for (uint16_t S = 0; S < W; ++S) {
+    iterator B = AVs.begin(), E = AVs.end();
+    // The registers in AVs are ordered according to the lexical order of
+    // the corresponding register cells. This means that the range of regis-
+    // ters in AVs that match a prefix of length L+1 will be contained in
+    // the range that matches a prefix of length L. This means that we can
+    // keep narrowing the search space as the prefix length goes up. This
+    // helps reduce the overall complexity of the search.
+    uint16_t L;
+    for (L = 0; L < W-S; ++L) {
+      // Compare against VR's bits starting at S, which emulates rotation
+      // of VR by S.
+      RegisterCellBitCompareSel RCB(VR, S+L, L, BVO, *CMS);
+      iterator NewB = std::lower_bound(B, E, VR, RCB);
+      iterator NewE = std::upper_bound(NewB, E, VR, RCB);
+      // For the registers that are eliminated from the next range, L is
+      // the longest prefix matching VR at position S (their prefixes
+      // differ from VR at S+L). If L>0, record this information for later
+      // use.
+      if (L > 0) {
+        for (iterator I = B; I != NewB; ++I)
+          LM[L].push_back(std::make_pair(*I, S));
+        for (iterator I = NewE; I != E; ++I)
+          LM[L].push_back(std::make_pair(*I, S));
+      }
+      B = NewB, E = NewE;
+      if (B == E)
+        break;
+    }
+    // Record the final register range. If this range is non-empty, then
+    // L=W-S.
+    assert(B == E || L == W-S);
+    if (B != E) {
+      for (iterator I = B; I != E; ++I)
+        LM[L].push_back(std::make_pair(*I, S));
+      // If B!=E, then we found a range of registers whose prefixes cover the
+      // rest of VR from position S. There is no need to further advance S.
+      break;
+    }
+  }
+
+  if (isDebug()) {
+    dbgs() << "Prefixes matching register " << PrintReg(VR, HRI) << "\n";
+    for (LRSMapType::iterator I = LM.begin(), E = LM.end(); I != E; ++I) {
+      dbgs() << "  L=" << I->first << ':';
+      const RSListType &LL = I->second;
+      for (unsigned i = 0, n = LL.size(); i < n; ++i)
+        dbgs() << " (" << PrintReg(LL[i].first, HRI) << ",@"
+               << LL[i].second << ')';
+      dbgs() << '\n';
+    }
+  }
+
+
+  bool Recorded = false;
+
+  for (iterator I = AVs.begin(), E = AVs.end(); I != E; ++I) {
+    unsigned SrcR = *I;
+    int FDi = -1, LDi = -1;   // First/last different bit.
+    const BitTracker::RegisterCell &AC = CMS->lookup(SrcR);
+    uint16_t AW = AC.width();
+    for (uint16_t i = 0, w = std::min(W, AW); i < w; ++i) {
+      if (RC[i] == AC[i])
+        continue;
+      if (FDi == -1)
+        FDi = i;
+      LDi = i;
+    }
+    if (FDi == -1)
+      continue;  // TODO (future): Record identical registers.
+    // Look for a register whose prefix could patch the range [FD..LD]
+    // where VR and SrcR differ.
+    uint16_t FD = FDi, LD = LDi;  // Switch to unsigned type.
+    uint16_t MinL = LD-FD+1;
+    for (uint16_t L = MinL; L < W; ++L) {
+      LRSMapType::iterator F = LM.find(L);
+      if (F == LM.end())
+        continue;
+      RSListType &LL = F->second;
+      for (unsigned i = 0, n = LL.size(); i < n; ++i) {
+        uint16_t S = LL[i].second;
+        // MinL is the minimum length of the prefix. Any length above MinL
+        // allows some flexibility as to where the prefix can start:
+        // given the extra length EL=L-MinL, the prefix must start between
+        // max(0,FD-EL) and FD.
+        if (S > FD)   // Starts too late.
+          continue;
+        uint16_t EL = L-MinL;
+        uint16_t LowS = (EL < FD) ? FD-EL : 0;
+        if (S < LowS) // Starts too early.
+          continue;
+        unsigned InsR = LL[i].first;
+        if (!isValidInsertForm(VR, SrcR, InsR, L, S))
+          continue;
+        if (isDebug()) {
+          dbgs() << PrintReg(VR, HRI) << " = insert(" << PrintReg(SrcR, HRI)
+                 << ',' << PrintReg(InsR, HRI) << ",#" << L << ",#"
+                 << S << ")\n";
+        }
+        IFRecordWithRegSet RR(IFRecord(SrcR, InsR, L, S), RegisterSet());
+        IFMap[VR].push_back(RR);
+        Recorded = true;
+      }
+    }
+  }
+
+  return Recorded;
+}
+
+
+void HexagonGenInsert::collectInBlock(MachineBasicBlock *B,
+      OrderedRegisterList &AVs) {
+  if (isDebug())
+    dbgs() << "visiting block BB#" << B->getNumber() << "\n";
+
+  // First, check if this block is reachable at all. If not, the bit tracker
+  // will not have any information about registers in it.
+  if (!CMS->BT.reached(B))
+    return;
+
+  bool DoConst = OptConst;
+  // Keep a separate set of registers defined in this block, so that we
+  // can remove them from the list of available registers once all DT
+  // successors have been processed.
+  RegisterSet BlockDefs, InsDefs;
+  for (MachineBasicBlock::iterator I = B->begin(), E = B->end(); I != E; ++I) {
+    MachineInstr *MI = &*I;
+    InsDefs.clear();
+    getInstrDefs(MI, InsDefs);
+    // Leave those alone. They are more transparent than "insert".
+    bool Skip = MI->isCopy() || MI->isRegSequence();
+
+    if (!Skip) {
+      // Visit all defined registers, and attempt to find the corresponding
+      // "insert" representations.
+      for (unsigned VR = InsDefs.find_first(); VR; VR = InsDefs.find_next(VR)) {
+        // Do not collect registers that are known to be compile-time cons-
+        // tants, unless requested.
+        if (!DoConst && isConstant(VR))
+          continue;
+        // If VR's cell contains a reference to VR, then VR cannot be defined
+        // via "insert". If VR is a constant that can be generated in a single
+        // instruction (without constant extenders), generating it via insert
+        // makes no sense.
+        if (findSelfReference(VR) || isSmallConstant(VR))
+          continue;
+
+        findRecordInsertForms(VR, AVs);
+      }
+    }
+
+    // Insert the defined registers into the list of available registers
+    // after they have been processed.
+    for (unsigned VR = InsDefs.find_first(); VR; VR = InsDefs.find_next(VR))
+      AVs.insert(VR);
+    BlockDefs.insert(InsDefs);
+  }
+
+  MachineDomTreeNode *N = MDT->getNode(B);
+  typedef GraphTraits<MachineDomTreeNode*> GTN;
+  typedef GTN::ChildIteratorType ChildIter;
+  for (ChildIter I = GTN::child_begin(N), E = GTN::child_end(N); I != E; ++I) {
+    MachineBasicBlock *SB = (*I)->getBlock();
+    collectInBlock(SB, AVs);
+  }
+
+  for (unsigned VR = BlockDefs.find_first(); VR; VR = BlockDefs.find_next(VR))
+    AVs.remove(VR);
+}
+
+
+void HexagonGenInsert::findRemovableRegisters(unsigned VR, IFRecord IF,
+      RegisterSet &RMs) const {
+  // For a given register VR and a insert form, find the registers that are
+  // used by the current definition of VR, and which would no longer be
+  // needed for it after the definition of VR is replaced with the insert
+  // form. These are the registers that could potentially become dead.
+  RegisterSet Regs[2];
+
+  unsigned S = 0;  // Register set selector.
+  Regs[S].insert(VR);
+
+  while (!Regs[S].empty()) {
+    // Breadth-first search.
+    unsigned OtherS = 1-S;
+    Regs[OtherS].clear();
+    for (unsigned R = Regs[S].find_first(); R; R = Regs[S].find_next(R)) {
+      Regs[S].remove(R);
+      if (R == IF.SrcR || R == IF.InsR)
+        continue;
+      // Check if a given register has bits that are references to any other
+      // registers. This is to detect situations where the instruction that
+      // defines register R takes register Q as an operand, but R itself does
+      // not contain any bits from Q. Loads are examples of how this could
+      // happen:
+      //   R = load Q
+      // In this case (assuming we do not have any knowledge about the loaded
+      // value), we must not treat R as a "conveyance" of the bits from Q.
+      // (The information in BT about R's bits would have them as constants,
+      // in case of zero-extending loads, or refs to R.)
+      if (!findNonSelfReference(R))
+        continue;
+      RMs.insert(R);
+      const MachineInstr *DefI = MRI->getVRegDef(R);
+      assert(DefI);
+      // Do not iterate past PHI nodes to avoid infinite loops. This can
+      // make the final set a bit less accurate, but the removable register
+      // sets are an approximation anyway.
+      if (DefI->isPHI())
+        continue;
+      getInstrUses(DefI, Regs[OtherS]);
+    }
+    S = OtherS;
+  }
+  // The register VR is added to the list as a side-effect of the algorithm,
+  // but it is not "potentially removable". A potentially removable register
+  // is one that may become unused (dead) after conversion to the insert form
+  // IF, and obviously VR (or its replacement) will not become dead by apply-
+  // ing IF.
+  RMs.remove(VR);
+}
+
+
+void HexagonGenInsert::computeRemovableRegisters() {
+  for (IFMapType::iterator I = IFMap.begin(), E = IFMap.end(); I != E; ++I) {
+    IFListType &LL = I->second;
+    for (unsigned i = 0, n = LL.size(); i < n; ++i)
+      findRemovableRegisters(I->first, LL[i].first, LL[i].second);
+  }
+}
+
+
+void HexagonGenInsert::pruneEmptyLists() {
+  // Remove all entries from the map, where the register has no insert forms
+  // associated with it.
+  typedef SmallVector<IFMapType::iterator,16> IterListType;
+  IterListType Prune;
+  for (IFMapType::iterator I = IFMap.begin(), E = IFMap.end(); I != E; ++I) {
+    if (I->second.size() == 0)
+      Prune.push_back(I);
+  }
+  for (unsigned i = 0, n = Prune.size(); i < n; ++i)
+    IFMap.erase(Prune[i]);
+}
+
+
+void HexagonGenInsert::pruneCoveredSets(unsigned VR) {
+  IFMapType::iterator F = IFMap.find(VR);
+  assert(F != IFMap.end());
+  IFListType &LL = F->second;
+
+  // First, examine the IF candidates for register VR whose removable-regis-
+  // ter sets are empty. This means that a given candidate will not help eli-
+  // minate any registers, but since "insert" is not a constant-extendable
+  // instruction, using such a candidate may reduce code size if the defini-
+  // tion of VR is constant-extended.
+  // If there exists a candidate with a non-empty set, the ones with empty
+  // sets will not be used and can be removed.
+  MachineInstr *DefVR = MRI->getVRegDef(VR);
+  bool DefEx = HII->isConstExtended(DefVR);
+  bool HasNE = false;
+  for (unsigned i = 0, n = LL.size(); i < n; ++i) {
+    if (LL[i].second.empty())
+      continue;
+    HasNE = true;
+    break;
+  }
+  if (!DefEx || HasNE) {
+    // The definition of VR is not constant-extended, or there is a candidate
+    // with a non-empty set. Remove all candidates with empty sets.
+    auto IsEmpty = [] (const IFRecordWithRegSet &IR) -> bool {
+      return IR.second.empty();
+    };
+    auto End = std::remove_if(LL.begin(), LL.end(), IsEmpty);
+    if (End != LL.end())
+      LL.erase(End, LL.end());
+  } else {
+    // The definition of VR is constant-extended, and all candidates have
+    // empty removable-register sets. Pick the maximum candidate, and remove
+    // all others. The "maximum" does not have any special meaning here, it
+    // is only so that the candidate that will remain on the list is selec-
+    // ted deterministically.
+    IFRecord MaxIF = LL[0].first;
+    for (unsigned i = 1, n = LL.size(); i < n; ++i) {
+      // If LL[MaxI] < LL[i], then MaxI = i.
+      const IFRecord &IF = LL[i].first;
+      unsigned M0 = BaseOrd[MaxIF.SrcR], M1 = BaseOrd[MaxIF.InsR];
+      unsigned R0 = BaseOrd[IF.SrcR], R1 = BaseOrd[IF.InsR];
+      if (M0 > R0)
+        continue;
+      if (M0 == R0) {
+        if (M1 > R1)
+          continue;
+        if (M1 == R1) {
+          if (MaxIF.Wdh > IF.Wdh)
+            continue;
+          if (MaxIF.Wdh == IF.Wdh && MaxIF.Off >= IF.Off)
+            continue;
+        }
+      }
+      // MaxIF < IF.
+      MaxIF = IF;
+    }
+    // Remove everything except the maximum candidate. All register sets
+    // are empty, so no need to preserve anything.
+    LL.clear();
+    LL.push_back(std::make_pair(MaxIF, RegisterSet()));
+  }
+
+  // Now, remove those whose sets of potentially removable registers are
+  // contained in another IF candidate for VR. For example, given these
+  // candidates for vreg45,
+  //   %vreg45:
+  //     (%vreg44,%vreg41,#9,#8), { %vreg42 }
+  //     (%vreg43,%vreg41,#9,#8), { %vreg42 %vreg44 }
+  // remove the first one, since it is contained in the second one.
+  for (unsigned i = 0, n = LL.size(); i < n; ) {
+    const RegisterSet &RMi = LL[i].second;
+    unsigned j = 0;
+    while (j < n) {
+      if (j != i && LL[j].second.includes(RMi))
+        break;
+      j++;
+    }
+    if (j == n) {   // RMi not contained in anything else.
+      i++;
+      continue;
+    }
+    LL.erase(LL.begin()+i);
+    n = LL.size();
+  }
+}
+
+
+void HexagonGenInsert::pruneUsesTooFar(unsigned VR, const UnsignedMap &RPO,
+      PairMapType &M) {
+  IFMapType::iterator F = IFMap.find(VR);
+  assert(F != IFMap.end());
+  IFListType &LL = F->second;
+  unsigned Cutoff = VRegDistCutoff;
+  const MachineInstr *DefV = MRI->getVRegDef(VR);
+
+  for (unsigned i = LL.size(); i > 0; --i) {
+    unsigned SR = LL[i-1].first.SrcR, IR = LL[i-1].first.InsR;
+    const MachineInstr *DefS = MRI->getVRegDef(SR);
+    const MachineInstr *DefI = MRI->getVRegDef(IR);
+    unsigned DSV = distance(DefS, DefV, RPO, M);
+    if (DSV < Cutoff) {
+      unsigned DIV = distance(DefI, DefV, RPO, M);
+      if (DIV < Cutoff)
+        continue;
+    }
+    LL.erase(LL.begin()+(i-1));
+  }
+}
+
+
+void HexagonGenInsert::pruneRegCopies(unsigned VR) {
+  IFMapType::iterator F = IFMap.find(VR);
+  assert(F != IFMap.end());
+  IFListType &LL = F->second;
+
+  auto IsCopy = [] (const IFRecordWithRegSet &IR) -> bool {
+    return IR.first.Wdh == 32 && (IR.first.Off == 0 || IR.first.Off == 32);
+  };
+  auto End = std::remove_if(LL.begin(), LL.end(), IsCopy);
+  if (End != LL.end())
+    LL.erase(End, LL.end());
+}
+
+
+void HexagonGenInsert::pruneCandidates() {
+  // Remove candidates that are not beneficial, regardless of the final
+  // selection method.
+  // First, remove candidates whose potentially removable set is a subset
+  // of another candidate's set.
+  for (IFMapType::iterator I = IFMap.begin(), E = IFMap.end(); I != E; ++I)
+    pruneCoveredSets(I->first);
+
+  UnsignedMap RPO;
+  typedef ReversePostOrderTraversal<const MachineFunction*> RPOTType;
+  RPOTType RPOT(MFN);
+  unsigned RPON = 0;
+  for (RPOTType::rpo_iterator I = RPOT.begin(), E = RPOT.end(); I != E; ++I)
+    RPO[(*I)->getNumber()] = RPON++;
+
+  PairMapType Memo; // Memoization map for distance calculation.
+  // Remove candidates that would use registers defined too far away.
+  for (IFMapType::iterator I = IFMap.begin(), E = IFMap.end(); I != E; ++I)
+    pruneUsesTooFar(I->first, RPO, Memo);
+
+  pruneEmptyLists();
+
+  for (IFMapType::iterator I = IFMap.begin(), E = IFMap.end(); I != E; ++I)
+    pruneRegCopies(I->first);
+}
+
+
+namespace {
+  // Class for comparing IF candidates for registers that have multiple of
+  // them. The smaller the candidate, according to this ordering, the better.
+  // First, compare the number of zeros in the associated potentially remova-
+  // ble register sets. "Zero" indicates that the register is very likely to
+  // become dead after this transformation.
+  // Second, compare "averages", i.e. use-count per size. The lower wins.
+  // After that, it does not really matter which one is smaller. Resolve
+  // the tie in some deterministic way.
+  struct IFOrdering {
+    IFOrdering(const UnsignedMap &UC, const RegisterOrdering &BO)
+      : UseC(UC), BaseOrd(BO) {}
+    bool operator() (const IFRecordWithRegSet &A,
+          const IFRecordWithRegSet &B) const;
+  private:
+    void stats(const RegisterSet &Rs, unsigned &Size, unsigned &Zero,
+          unsigned &Sum) const;
+    const UnsignedMap &UseC;
+    const RegisterOrdering &BaseOrd;
+  };
+}
+
+
+bool IFOrdering::operator() (const IFRecordWithRegSet &A,
+      const IFRecordWithRegSet &B) const {
+  unsigned SizeA = 0, ZeroA = 0, SumA = 0;
+  unsigned SizeB = 0, ZeroB = 0, SumB = 0;
+  stats(A.second, SizeA, ZeroA, SumA);
+  stats(B.second, SizeB, ZeroB, SumB);
+
+  // We will pick the minimum element. The more zeros, the better.
+  if (ZeroA != ZeroB)
+    return ZeroA > ZeroB;
+  // Compare SumA/SizeA with SumB/SizeB, lower is better.
+  uint64_t AvgA = SumA*SizeB, AvgB = SumB*SizeA;
+  if (AvgA != AvgB)
+    return AvgA < AvgB;
+
+  // The sets compare identical so far. Resort to comparing the IF records.
+  // The actual values don't matter, this is only for determinism.
+  unsigned OSA = BaseOrd[A.first.SrcR], OSB = BaseOrd[B.first.SrcR];
+  if (OSA != OSB)
+    return OSA < OSB;
+  unsigned OIA = BaseOrd[A.first.InsR], OIB = BaseOrd[B.first.InsR];
+  if (OIA != OIB)
+    return OIA < OIB;
+  if (A.first.Wdh != B.first.Wdh)
+    return A.first.Wdh < B.first.Wdh;
+  return A.first.Off < B.first.Off;
+}
+
+
+void IFOrdering::stats(const RegisterSet &Rs, unsigned &Size, unsigned &Zero,
+      unsigned &Sum) const {
+  for (unsigned R = Rs.find_first(); R; R = Rs.find_next(R)) {
+    UnsignedMap::const_iterator F = UseC.find(R);
+    assert(F != UseC.end());
+    unsigned UC = F->second;
+    if (UC == 0)
+      Zero++;
+    Sum += UC;
+    Size++;
+  }
+}
+
+
+void HexagonGenInsert::selectCandidates() {
+  // Some registers may have multiple valid candidates. Pick the best one
+  // (or decide not to use any).
+
+  // Compute the "removability" measure of R:
+  // For each potentially removable register R, record the number of regis-
+  // ters with IF candidates, where R appears in at least one set.
+  RegisterSet AllRMs;
+  UnsignedMap UseC, RemC;
+  IFMapType::iterator End = IFMap.end();
+
+  for (IFMapType::iterator I = IFMap.begin(); I != End; ++I) {
+    const IFListType &LL = I->second;
+    RegisterSet TT;
+    for (unsigned i = 0, n = LL.size(); i < n; ++i)
+      TT.insert(LL[i].second);
+    for (unsigned R = TT.find_first(); R; R = TT.find_next(R))
+      RemC[R]++;
+    AllRMs.insert(TT);
+  }
+
+  for (unsigned R = AllRMs.find_first(); R; R = AllRMs.find_next(R)) {
+    typedef MachineRegisterInfo::use_nodbg_iterator use_iterator;
+    typedef SmallSet<const MachineInstr*,16> InstrSet;
+    InstrSet UIs;
+    // Count as the number of instructions in which R is used, not the
+    // number of operands.
+    use_iterator E = MRI->use_nodbg_end();
+    for (use_iterator I = MRI->use_nodbg_begin(R); I != E; ++I)
+      UIs.insert(I->getParent());
+    unsigned C = UIs.size();
+    // Calculate a measure, which is the number of instructions using R,
+    // minus the "removability" count computed earlier.
+    unsigned D = RemC[R];
+    UseC[R] = (C > D) ? C-D : 0;  // doz
+  }
+
+
+  bool SelectAll0 = OptSelectAll0, SelectHas0 = OptSelectHas0;
+  if (!SelectAll0 && !SelectHas0)
+    SelectAll0 = true;
+
+  // The smaller the number UseC for a given register R, the "less used"
+  // R is aside from the opportunities for removal offered by generating
+  // "insert" instructions.
+  // Iterate over the IF map, and for those registers that have multiple
+  // candidates, pick the minimum one according to IFOrdering.
+  IFOrdering IFO(UseC, BaseOrd);
+  for (IFMapType::iterator I = IFMap.begin(); I != End; ++I) {
+    IFListType &LL = I->second;
+    if (LL.empty())
+      continue;
+    // Get the minimum element, remember it and clear the list. If the
+    // element found is adequate, we will put it back on the list, other-
+    // wise the list will remain empty, and the entry for this register
+    // will be removed (i.e. this register will not be replaced by insert).
+    IFListType::iterator MinI = std::min_element(LL.begin(), LL.end(), IFO);
+    assert(MinI != LL.end());
+    IFRecordWithRegSet M = *MinI;
+    LL.clear();
+
+    // We want to make sure that this replacement will have a chance to be
+    // beneficial, and that means that we want to have indication that some
+    // register will be removed. The most likely registers to be eliminated
+    // are the use operands in the definition of I->first. Accept/reject a
+    // candidate based on how many of its uses it can potentially eliminate.
+
+    RegisterSet Us;
+    const MachineInstr *DefI = MRI->getVRegDef(I->first);
+    getInstrUses(DefI, Us);
+    bool Accept = false;
+
+    if (SelectAll0) {
+      bool All0 = true;
+      for (unsigned R = Us.find_first(); R; R = Us.find_next(R)) {
+        if (UseC[R] == 0)
+          continue;
+        All0 = false;
+        break;
+      }
+      Accept = All0;
+    } else if (SelectHas0) {
+      bool Has0 = false;
+      for (unsigned R = Us.find_first(); R; R = Us.find_next(R)) {
+        if (UseC[R] != 0)
+          continue;
+        Has0 = true;
+        break;
+      }
+      Accept = Has0;
+    }
+    if (Accept)
+      LL.push_back(M);
+  }
+
+  // Remove candidates that add uses of removable registers, unless the
+  // removable registers are among replacement candidates.
+  // Recompute the removable registers, since some candidates may have
+  // been eliminated.
+  AllRMs.clear();
+  for (IFMapType::iterator I = IFMap.begin(); I != End; ++I) {
+    const IFListType &LL = I->second;
+    if (LL.size() > 0)
+      AllRMs.insert(LL[0].second);
+  }
+  for (IFMapType::iterator I = IFMap.begin(); I != End; ++I) {
+    IFListType &LL = I->second;
+    if (LL.size() == 0)
+      continue;
+    unsigned SR = LL[0].first.SrcR, IR = LL[0].first.InsR;
+    if (AllRMs[SR] || AllRMs[IR])
+      LL.clear();
+  }
+
+  pruneEmptyLists();
+}
+
+
+bool HexagonGenInsert::generateInserts() {
+  // Create a new register for each one from IFMap, and store them in the
+  // map.
+  UnsignedMap RegMap;
+  for (IFMapType::iterator I = IFMap.begin(), E = IFMap.end(); I != E; ++I) {
+    unsigned VR = I->first;
+    const TargetRegisterClass *RC = MRI->getRegClass(VR);
+    unsigned NewVR = MRI->createVirtualRegister(RC);
+    RegMap[VR] = NewVR;
+  }
+
+  // We can generate the "insert" instructions using potentially stale re-
+  // gisters: SrcR and InsR for a given VR may be among other registers that
+  // are also replaced. This is fine, we will do the mass "rauw" a bit later.
+  for (IFMapType::iterator I = IFMap.begin(), E = IFMap.end(); I != E; ++I) {
+    MachineInstr *MI = MRI->getVRegDef(I->first);
+    MachineBasicBlock &B = *MI->getParent();
+    DebugLoc DL = MI->getDebugLoc();
+    unsigned NewR = RegMap[I->first];
+    bool R32 = MRI->getRegClass(NewR) == &Hexagon::IntRegsRegClass;
+    const MCInstrDesc &D = R32 ? HII->get(Hexagon::S2_insert)
+                               : HII->get(Hexagon::S2_insertp);
+    IFRecord IF = I->second[0].first;
+    unsigned Wdh = IF.Wdh, Off = IF.Off;
+    unsigned InsS = 0;
+    if (R32 && MRI->getRegClass(IF.InsR) == &Hexagon::DoubleRegsRegClass) {
+      InsS = Hexagon::subreg_loreg;
+      if (Off >= 32) {
+        InsS = Hexagon::subreg_hireg;
+        Off -= 32;
+      }
+    }
+    // Advance to the proper location for inserting instructions. This could
+    // be B.end().
+    MachineBasicBlock::iterator At = MI;
+    if (MI->isPHI())
+      At = B.getFirstNonPHI();
+
+    BuildMI(B, At, DL, D, NewR)
+      .addReg(IF.SrcR)
+      .addReg(IF.InsR, 0, InsS)
+      .addImm(Wdh)
+      .addImm(Off);
+
+    MRI->clearKillFlags(IF.SrcR);
+    MRI->clearKillFlags(IF.InsR);
+  }
+
+  for (IFMapType::iterator I = IFMap.begin(), E = IFMap.end(); I != E; ++I) {
+    MachineInstr *DefI = MRI->getVRegDef(I->first);
+    MRI->replaceRegWith(I->first, RegMap[I->first]);
+    DefI->eraseFromParent();
+  }
+
+  return true;
+}
+
+
+bool HexagonGenInsert::removeDeadCode(MachineDomTreeNode *N) {
+  bool Changed = false;
+  typedef GraphTraits<MachineDomTreeNode*> GTN;
+  for (auto I = GTN::child_begin(N), E = GTN::child_end(N); I != E; ++I)
+    Changed |= removeDeadCode(*I);
+
+  MachineBasicBlock *B = N->getBlock();
+  std::vector<MachineInstr*> Instrs;
+  for (auto I = B->rbegin(), E = B->rend(); I != E; ++I)
+    Instrs.push_back(&*I);
+
+  for (auto I = Instrs.begin(), E = Instrs.end(); I != E; ++I) {
+    MachineInstr *MI = *I;
+    unsigned Opc = MI->getOpcode();
+    // Do not touch lifetime markers. This is why the target-independent DCE
+    // cannot be used.
+    if (Opc == TargetOpcode::LIFETIME_START ||
+        Opc == TargetOpcode::LIFETIME_END)
+      continue;
+    bool Store = false;
+    if (MI->isInlineAsm() || !MI->isSafeToMove(nullptr, Store))
+      continue;
+
+    bool AllDead = true;
+    SmallVector<unsigned,2> Regs;
+    for (ConstMIOperands Op(MI); Op.isValid(); ++Op) {
+      if (!Op->isReg() || !Op->isDef())
+        continue;
+      unsigned R = Op->getReg();
+      if (!TargetRegisterInfo::isVirtualRegister(R) ||
+          !MRI->use_nodbg_empty(R)) {
+        AllDead = false;
+        break;
+      }
+      Regs.push_back(R);
+    }
+    if (!AllDead)
+      continue;
+
+    B->erase(MI);
+    for (unsigned I = 0, N = Regs.size(); I != N; ++I)
+      MRI->markUsesInDebugValueAsUndef(Regs[I]);
+    Changed = true;
+  }
+
+  return Changed;
+}
+
+
+bool HexagonGenInsert::runOnMachineFunction(MachineFunction &MF) {
+  bool Timing = OptTiming, TimingDetail = Timing && OptTimingDetail;
+  bool Changed = false;
+  TimerGroup __G("hexinsert");
+  NamedRegionTimer __T("hexinsert", Timing && !TimingDetail);
+
+  // Sanity check: one, but not both.
+  assert(!OptSelectAll0 || !OptSelectHas0);
+
+  IFMap.clear();
+  BaseOrd.clear();
+  CellOrd.clear();
+
+  const auto &ST = MF.getSubtarget<HexagonSubtarget>();
+  HII = ST.getInstrInfo();
+  HRI = ST.getRegisterInfo();
+  MFN = &MF;
+  MRI = &MF.getRegInfo();
+  MDT = &getAnalysis<MachineDominatorTree>();
+
+  // Clean up before any further processing, so that dead code does not
+  // get used in a newly generated "insert" instruction. Have a custom
+  // version of DCE that preserves lifetime markers. Without it, merging
+  // of stack objects can fail to recognize and merge disjoint objects
+  // leading to unnecessary stack growth.
+  Changed |= removeDeadCode(MDT->getRootNode());
+
+  const HexagonEvaluator HE(*HRI, *MRI, *HII, MF);
+  BitTracker BTLoc(HE, MF);
+  BTLoc.trace(isDebug());
+  BTLoc.run();
+  CellMapShadow MS(BTLoc);
+  CMS = &MS;
+
+  buildOrderingMF(BaseOrd);
+  buildOrderingBT(BaseOrd, CellOrd);
+
+  if (isDebug()) {
+    dbgs() << "Cell ordering:\n";
+    for (RegisterOrdering::iterator I = CellOrd.begin(), E = CellOrd.end();
+        I != E; ++I) {
+      unsigned VR = I->first, Pos = I->second;
+      dbgs() << PrintReg(VR, HRI) << " -> " << Pos << "\n";
+    }
+  }
+
+  // Collect candidates for conversion into the insert forms.
+  MachineBasicBlock *RootB = MDT->getRoot();
+  OrderedRegisterList AvailR(CellOrd);
+
+  {
+    NamedRegionTimer _T("collection", "hexinsert", TimingDetail);
+    collectInBlock(RootB, AvailR);
+    // Complete the information gathered in IFMap.
+    computeRemovableRegisters();
+  }
+
+  if (isDebug()) {
+    dbgs() << "Candidates after collection:\n";
+    dump_map();
+  }
+
+  if (IFMap.empty())
+    return false;
+
+  {
+    NamedRegionTimer _T("pruning", "hexinsert", TimingDetail);
+    pruneCandidates();
+  }
+
+  if (isDebug()) {
+    dbgs() << "Candidates after pruning:\n";
+    dump_map();
+  }
+
+  if (IFMap.empty())
+    return false;
+
+  {
+    NamedRegionTimer _T("selection", "hexinsert", TimingDetail);
+    selectCandidates();
+  }
+
+  if (isDebug()) {
+    dbgs() << "Candidates after selection:\n";
+    dump_map();
+  }
+
+  // Filter out vregs beyond the cutoff.
+  if (VRegIndexCutoff.getPosition()) {
+    unsigned Cutoff = VRegIndexCutoff;
+    typedef SmallVector<IFMapType::iterator,16> IterListType;
+    IterListType Out;
+    for (IFMapType::iterator I = IFMap.begin(), E = IFMap.end(); I != E; ++I) {
+      unsigned Idx = TargetRegisterInfo::virtReg2Index(I->first);
+      if (Idx >= Cutoff)
+        Out.push_back(I);
+    }
+    for (unsigned i = 0, n = Out.size(); i < n; ++i)
+      IFMap.erase(Out[i]);
+  }
+
+  {
+    NamedRegionTimer _T("generation", "hexinsert", TimingDetail);
+    Changed = generateInserts();
+  }
+
+  return Changed;
+}
+
+
+FunctionPass *llvm::createHexagonGenInsert() {
+  return new HexagonGenInsert();
+}
+
+
+//===----------------------------------------------------------------------===//
+//                         Public Constructor Functions
+//===----------------------------------------------------------------------===//
+
+INITIALIZE_PASS_BEGIN(HexagonGenInsert, "hexinsert",
+  "Hexagon generate \"insert\" instructions", false, false)
+INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
+INITIALIZE_PASS_END(HexagonGenInsert, "hexinsert",
+  "Hexagon generate \"insert\" instructions", false, false)

Modified: llvm/trunk/lib/Target/Hexagon/HexagonTargetMachine.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/Hexagon/HexagonTargetMachine.cpp?rev=241683&r1=241682&r2=241683&view=diff
==============================================================================
--- llvm/trunk/lib/Target/Hexagon/HexagonTargetMachine.cpp (original)
+++ llvm/trunk/lib/Target/Hexagon/HexagonTargetMachine.cpp Wed Jul  8 09:47:34 2015
@@ -37,6 +37,8 @@ static cl::opt<bool> EnableExpandCondset
   cl::init(true), cl::Hidden, cl::ZeroOrMore,
   cl::desc("Early expansion of MUX"));
 
+static cl::opt<bool> EnableGenInsert("hexagon-insert", cl::init(true),
+  cl::Hidden, cl::desc("Generate \"insert\" instructions"));
 
 /// HexagonTargetMachineModule - Note that this is used on hosts that
 /// cannot link in a library unless there are references into the
@@ -64,12 +66,12 @@ namespace llvm {
   FunctionPass *createHexagonISelDag(HexagonTargetMachine &TM,
                                      CodeGenOpt::Level OptLevel);
   FunctionPass *createHexagonDelaySlotFillerPass(const TargetMachine &TM);
-  FunctionPass *createHexagonFPMoverPass(const TargetMachine &TM);
   FunctionPass *createHexagonRemoveExtendArgs(const HexagonTargetMachine &TM);
   FunctionPass *createHexagonCFGOptimizer();
 
   FunctionPass *createHexagonSplitConst32AndConst64();
   FunctionPass *createHexagonExpandPredSpillCode();
+  FunctionPass *createHexagonGenInsert();
   FunctionPass *createHexagonHardwareLoops();
   FunctionPass *createHexagonPeephole();
   FunctionPass *createHexagonFixupHwLoops();
@@ -146,6 +148,8 @@ bool HexagonPassConfig::addInstSelector(
   if (!NoOpt) {
     addPass(createHexagonPeephole());
     printAndVerify("After hexagon peephole pass");
+    if (EnableGenInsert)
+      addPass(createHexagonGenInsert(), false);
   }
 
   return false;

Added: llvm/trunk/test/CodeGen/Hexagon/insert-basic.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/Hexagon/insert-basic.ll?rev=241683&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/Hexagon/insert-basic.ll (added)
+++ llvm/trunk/test/CodeGen/Hexagon/insert-basic.ll Wed Jul  8 09:47:34 2015
@@ -0,0 +1,66 @@
+; RUN: llc -O2 -march=hexagon < %s | FileCheck %s
+; CHECK-DAG: insert(r{{[0-9]*}}, #17, #0)
+; CHECK-DAG: insert(r{{[0-9]*}}, #18, #0)
+; CHECK-DAG: insert(r{{[0-9]*}}, #22, #0)
+; CHECK-DAG: insert(r{{[0-9]*}}, #12, #0)
+
+; C source:
+; typedef struct {
+;   unsigned x1:23;
+;   unsigned x2:17;
+;   unsigned x3:18;
+;   unsigned x4:22;
+;   unsigned x5:12;
+; } structx_t;
+;
+; void foo(structx_t *px, int y1, int y2, int y3, int y4, int y5) {
+;   px->x1 = y1;
+;   px->x2 = y2;
+;   px->x3 = y3;
+;   px->x4 = y4;
+;   px->x5 = y5;
+; }
+
+target datalayout = "e-p:32:32:32-i64:64:64-i32:32:32-i16:16:16-i1:32:32-f64:64:64-f32:32:32-v64:64:64-v32:32:32-a0:0-n16:32"
+target triple = "hexagon"
+
+%struct.structx_t = type { [3 x i8], i8, [3 x i8], i8, [3 x i8], i8, [3 x i8], i8, [2 x i8], [2 x i8] }
+
+define void @foo(%struct.structx_t* nocapture %px, i32 %y1, i32 %y2, i32 %y3, i32 %y4, i32 %y5) nounwind {
+entry:
+  %bf.value = and i32 %y1, 8388607
+  %0 = bitcast %struct.structx_t* %px to i32*
+  %1 = load i32, i32* %0, align 4
+  %2 = and i32 %1, -8388608
+  %3 = or i32 %2, %bf.value
+  store i32 %3, i32* %0, align 4
+  %bf.value1 = and i32 %y2, 131071
+  %bf.field.offs = getelementptr %struct.structx_t, %struct.structx_t* %px, i32 0, i32 0, i32 4
+  %4 = bitcast i8* %bf.field.offs to i32*
+  %5 = load i32, i32* %4, align 4
+  %6 = and i32 %5, -131072
+  %7 = or i32 %6, %bf.value1
+  store i32 %7, i32* %4, align 4
+  %bf.value2 = and i32 %y3, 262143
+  %bf.field.offs3 = getelementptr %struct.structx_t, %struct.structx_t* %px, i32 0, i32 0, i32 8
+  %8 = bitcast i8* %bf.field.offs3 to i32*
+  %9 = load i32, i32* %8, align 4
+  %10 = and i32 %9, -262144
+  %11 = or i32 %10, %bf.value2
+  store i32 %11, i32* %8, align 4
+  %bf.value4 = and i32 %y4, 4194303
+  %bf.field.offs5 = getelementptr %struct.structx_t, %struct.structx_t* %px, i32 0, i32 0, i32 12
+  %12 = bitcast i8* %bf.field.offs5 to i32*
+  %13 = load i32, i32* %12, align 4
+  %14 = and i32 %13, -4194304
+  %15 = or i32 %14, %bf.value4
+  store i32 %15, i32* %12, align 4
+  %bf.value6 = and i32 %y5, 4095
+  %bf.field.offs7 = getelementptr %struct.structx_t, %struct.structx_t* %px, i32 0, i32 0, i32 16
+  %16 = bitcast i8* %bf.field.offs7 to i32*
+  %17 = load i32, i32* %16, align 4
+  %18 = and i32 %17, -4096
+  %19 = or i32 %18, %bf.value6
+  store i32 %19, i32* %16, align 4
+  ret void
+}





More information about the llvm-commits mailing list