Home | History | Annotate | Download | only in Hexagon
      1 //===----- HexagonPacketizer.cpp - vliw packetizer ---------------------===//
      2 //
      3 //                     The LLVM Compiler Infrastructure
      4 //
      5 // This file is distributed under the University of Illinois Open Source
      6 // License. See LICENSE.TXT for details.
      7 //
      8 //===----------------------------------------------------------------------===//
      9 //
     10 // This implements a simple VLIW packetizer using DFA. The packetizer works on
     11 // machine basic blocks. For each instruction I in BB, the packetizer consults
     12 // the DFA to see if machine resources are available to execute I. If so, the
     13 // packetizer checks if I depends on any instruction J in the current packet.
     14 // If no dependency is found, I is added to current packet and machine resource
     15 // is marked as taken. If any dependency is found, a target API call is made to
     16 // prune the dependence.
     17 //
     18 //===----------------------------------------------------------------------===//
     19 #define DEBUG_TYPE "packets"
     20 #include "llvm/CodeGen/DFAPacketizer.h"
     21 #include "llvm/CodeGen/Passes.h"
     22 #include "llvm/CodeGen/MachineDominators.h"
     23 #include "llvm/CodeGen/MachineFunctionPass.h"
     24 #include "llvm/CodeGen/MachineLoopInfo.h"
     25 #include "llvm/CodeGen/ScheduleDAG.h"
     26 #include "llvm/CodeGen/ScheduleDAGInstrs.h"
     27 #include "llvm/CodeGen/LatencyPriorityQueue.h"
     28 #include "llvm/CodeGen/SchedulerRegistry.h"
     29 #include "llvm/CodeGen/MachineFrameInfo.h"
     30 #include "llvm/CodeGen/MachineInstrBuilder.h"
     31 #include "llvm/CodeGen/MachineRegisterInfo.h"
     32 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
     33 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
     34 #include "llvm/Target/TargetMachine.h"
     35 #include "llvm/Target/TargetInstrInfo.h"
     36 #include "llvm/Target/TargetRegisterInfo.h"
     37 #include "llvm/ADT/DenseMap.h"
     38 #include "llvm/ADT/Statistic.h"
     39 #include "llvm/Support/MathExtras.h"
     40 #include "llvm/MC/MCInstrItineraries.h"
     41 #include "llvm/Support/Compiler.h"
     42 #include "llvm/Support/CommandLine.h"
     43 #include "llvm/Support/Debug.h"
     44 #include "Hexagon.h"
     45 #include "HexagonTargetMachine.h"
     46 #include "HexagonRegisterInfo.h"
     47 #include "HexagonSubtarget.h"
     48 #include "HexagonMachineFunctionInfo.h"
     49 
     50 #include <map>
     51 #include <vector>
     52 
     53 using namespace llvm;
     54 
     55 static cl::opt<bool> PacketizeVolatiles("hexagon-packetize-volatiles",
     56       cl::ZeroOrMore, cl::Hidden, cl::init(true),
     57       cl::desc("Allow non-solo packetization of volatile memory references"));
     58 
     59 namespace llvm {
     60   void initializeHexagonPacketizerPass(PassRegistry&);
     61 }
     62 
     63 
     64 namespace {
     65   class HexagonPacketizer : public MachineFunctionPass {
     66 
     67   public:
     68     static char ID;
     69     HexagonPacketizer() : MachineFunctionPass(ID) {
     70       initializeHexagonPacketizerPass(*PassRegistry::getPassRegistry());
     71     }
     72 
     73     void getAnalysisUsage(AnalysisUsage &AU) const {
     74       AU.setPreservesCFG();
     75       AU.addRequired<MachineDominatorTree>();
     76       AU.addRequired<MachineBranchProbabilityInfo>();
     77       AU.addPreserved<MachineDominatorTree>();
     78       AU.addRequired<MachineLoopInfo>();
     79       AU.addPreserved<MachineLoopInfo>();
     80       MachineFunctionPass::getAnalysisUsage(AU);
     81     }
     82 
     83     const char *getPassName() const {
     84       return "Hexagon Packetizer";
     85     }
     86 
     87     bool runOnMachineFunction(MachineFunction &Fn);
     88   };
     89   char HexagonPacketizer::ID = 0;
     90 
     91   class HexagonPacketizerList : public VLIWPacketizerList {
     92 
     93   private:
     94 
     95     // Has the instruction been promoted to a dot-new instruction.
     96     bool PromotedToDotNew;
     97 
     98     // Has the instruction been glued to allocframe.
     99     bool GlueAllocframeStore;
    100 
    101     // Has the feeder instruction been glued to new value jump.
    102     bool GlueToNewValueJump;
    103 
    104     // Check if there is a dependence between some instruction already in this
    105     // packet and this instruction.
    106     bool Dependence;
    107 
    108     // Only check for dependence if there are resources available to
    109     // schedule this instruction.
    110     bool FoundSequentialDependence;
    111 
    112     /// \brief A handle to the branch probability pass.
    113    const MachineBranchProbabilityInfo *MBPI;
    114 
    115    // Track MIs with ignored dependece.
    116    std::vector<MachineInstr*> IgnoreDepMIs;
    117 
    118   public:
    119     // Ctor.
    120     HexagonPacketizerList(MachineFunction &MF, MachineLoopInfo &MLI,
    121                           MachineDominatorTree &MDT,
    122                           const MachineBranchProbabilityInfo *MBPI);
    123 
    124     // initPacketizerState - initialize some internal flags.
    125     void initPacketizerState();
    126 
    127     // ignorePseudoInstruction - Ignore bundling of pseudo instructions.
    128     bool ignorePseudoInstruction(MachineInstr *MI, MachineBasicBlock *MBB);
    129 
    130     // isSoloInstruction - return true if instruction MI can not be packetized
    131     // with any other instruction, which means that MI itself is a packet.
    132     bool isSoloInstruction(MachineInstr *MI);
    133 
    134     // isLegalToPacketizeTogether - Is it legal to packetize SUI and SUJ
    135     // together.
    136     bool isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ);
    137 
    138     // isLegalToPruneDependencies - Is it legal to prune dependece between SUI
    139     // and SUJ.
    140     bool isLegalToPruneDependencies(SUnit *SUI, SUnit *SUJ);
    141 
    142     MachineBasicBlock::iterator addToPacket(MachineInstr *MI);
    143   private:
    144     bool IsCallDependent(MachineInstr* MI, SDep::Kind DepType, unsigned DepReg);
    145     bool PromoteToDotNew(MachineInstr* MI, SDep::Kind DepType,
    146                          MachineBasicBlock::iterator &MII,
    147                          const TargetRegisterClass* RC);
    148     bool CanPromoteToDotNew(MachineInstr* MI, SUnit* PacketSU,
    149                             unsigned DepReg,
    150                             std::map <MachineInstr*, SUnit*> MIToSUnit,
    151                             MachineBasicBlock::iterator &MII,
    152                             const TargetRegisterClass* RC);
    153     bool CanPromoteToNewValue(MachineInstr* MI, SUnit* PacketSU,
    154                               unsigned DepReg,
    155                               std::map <MachineInstr*, SUnit*> MIToSUnit,
    156                               MachineBasicBlock::iterator &MII);
    157     bool CanPromoteToNewValueStore(MachineInstr* MI, MachineInstr* PacketMI,
    158                                    unsigned DepReg,
    159                                    std::map <MachineInstr*, SUnit*> MIToSUnit);
    160     bool DemoteToDotOld(MachineInstr* MI);
    161     bool ArePredicatesComplements(MachineInstr* MI1, MachineInstr* MI2,
    162                     std::map <MachineInstr*, SUnit*> MIToSUnit);
    163     bool RestrictingDepExistInPacket(MachineInstr*,
    164                     unsigned, std::map <MachineInstr*, SUnit*>);
    165     bool isNewifiable(MachineInstr* MI);
    166     bool isCondInst(MachineInstr* MI);
    167     bool tryAllocateResourcesForConstExt(MachineInstr* MI);
    168     bool canReserveResourcesForConstExt(MachineInstr *MI);
    169     void reserveResourcesForConstExt(MachineInstr* MI);
    170     bool isNewValueInst(MachineInstr* MI);
    171   };
    172 }
    173 
    174 INITIALIZE_PASS_BEGIN(HexagonPacketizer, "packets", "Hexagon Packetizer",
    175                       false, false)
    176 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
    177 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
    178 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
    179 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
    180 INITIALIZE_PASS_END(HexagonPacketizer, "packets", "Hexagon Packetizer",
    181                     false, false)
    182 
    183 
    184 // HexagonPacketizerList Ctor.
    185 HexagonPacketizerList::HexagonPacketizerList(
    186   MachineFunction &MF, MachineLoopInfo &MLI,MachineDominatorTree &MDT,
    187   const MachineBranchProbabilityInfo *MBPI)
    188   : VLIWPacketizerList(MF, MLI, MDT, true){
    189   this->MBPI = MBPI;
    190 }
    191 
    192 bool HexagonPacketizer::runOnMachineFunction(MachineFunction &Fn) {
    193   const TargetInstrInfo *TII = Fn.getTarget().getInstrInfo();
    194   MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
    195   MachineDominatorTree &MDT = getAnalysis<MachineDominatorTree>();
    196   const MachineBranchProbabilityInfo *MBPI =
    197     &getAnalysis<MachineBranchProbabilityInfo>();
    198   // Instantiate the packetizer.
    199   HexagonPacketizerList Packetizer(Fn, MLI, MDT, MBPI);
    200 
    201   // DFA state table should not be empty.
    202   assert(Packetizer.getResourceTracker() && "Empty DFA table!");
    203 
    204   //
    205   // Loop over all basic blocks and remove KILL pseudo-instructions
    206   // These instructions confuse the dependence analysis. Consider:
    207   // D0 = ...   (Insn 0)
    208   // R0 = KILL R0, D0 (Insn 1)
    209   // R0 = ... (Insn 2)
    210   // Here, Insn 1 will result in the dependence graph not emitting an output
    211   // dependence between Insn 0 and Insn 2. This can lead to incorrect
    212   // packetization
    213   //
    214   for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
    215        MBB != MBBe; ++MBB) {
    216     MachineBasicBlock::iterator End = MBB->end();
    217     MachineBasicBlock::iterator MI = MBB->begin();
    218     while (MI != End) {
    219       if (MI->isKill()) {
    220         MachineBasicBlock::iterator DeleteMI = MI;
    221         ++MI;
    222         MBB->erase(DeleteMI);
    223         End = MBB->end();
    224         continue;
    225       }
    226       ++MI;
    227     }
    228   }
    229 
    230   // Loop over all of the basic blocks.
    231   for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
    232        MBB != MBBe; ++MBB) {
    233     // Find scheduling regions and schedule / packetize each region.
    234     unsigned RemainingCount = MBB->size();
    235     for(MachineBasicBlock::iterator RegionEnd = MBB->end();
    236         RegionEnd != MBB->begin();) {
    237       // The next region starts above the previous region. Look backward in the
    238       // instruction stream until we find the nearest boundary.
    239       MachineBasicBlock::iterator I = RegionEnd;
    240       for(;I != MBB->begin(); --I, --RemainingCount) {
    241         if (TII->isSchedulingBoundary(llvm::prior(I), MBB, Fn))
    242           break;
    243       }
    244       I = MBB->begin();
    245 
    246       // Skip empty scheduling regions.
    247       if (I == RegionEnd) {
    248         RegionEnd = llvm::prior(RegionEnd);
    249         --RemainingCount;
    250         continue;
    251       }
    252       // Skip regions with one instruction.
    253       if (I == llvm::prior(RegionEnd)) {
    254         RegionEnd = llvm::prior(RegionEnd);
    255         continue;
    256       }
    257 
    258       Packetizer.PacketizeMIs(MBB, I, RegionEnd);
    259       RegionEnd = I;
    260     }
    261   }
    262 
    263   return true;
    264 }
    265 
    266 
    267 static bool IsIndirectCall(MachineInstr* MI) {
    268   return ((MI->getOpcode() == Hexagon::CALLR) ||
    269           (MI->getOpcode() == Hexagon::CALLRv3));
    270 }
    271 
    272 // Reserve resources for constant extender. Trigure an assertion if
    273 // reservation fail.
    274 void HexagonPacketizerList::reserveResourcesForConstExt(MachineInstr* MI) {
    275   const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
    276   MachineFunction *MF = MI->getParent()->getParent();
    277   MachineInstr *PseudoMI = MF->CreateMachineInstr(QII->get(Hexagon::IMMEXT_i),
    278                                                   MI->getDebugLoc());
    279 
    280   if (ResourceTracker->canReserveResources(PseudoMI)) {
    281     ResourceTracker->reserveResources(PseudoMI);
    282     MI->getParent()->getParent()->DeleteMachineInstr(PseudoMI);
    283   } else {
    284     MI->getParent()->getParent()->DeleteMachineInstr(PseudoMI);
    285     llvm_unreachable("can not reserve resources for constant extender.");
    286   }
    287   return;
    288 }
    289 
    290 bool HexagonPacketizerList::canReserveResourcesForConstExt(MachineInstr *MI) {
    291   const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
    292   assert((QII->isExtended(MI) || QII->isConstExtended(MI)) &&
    293          "Should only be called for constant extended instructions");
    294   MachineFunction *MF = MI->getParent()->getParent();
    295   MachineInstr *PseudoMI = MF->CreateMachineInstr(QII->get(Hexagon::IMMEXT_i),
    296                                                   MI->getDebugLoc());
    297   bool CanReserve = ResourceTracker->canReserveResources(PseudoMI);
    298   MF->DeleteMachineInstr(PseudoMI);
    299   return CanReserve;
    300 }
    301 
    302 // Allocate resources (i.e. 4 bytes) for constant extender. If succeed, return
    303 // true, otherwise, return false.
    304 bool HexagonPacketizerList::tryAllocateResourcesForConstExt(MachineInstr* MI) {
    305   const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
    306   MachineFunction *MF = MI->getParent()->getParent();
    307   MachineInstr *PseudoMI = MF->CreateMachineInstr(QII->get(Hexagon::IMMEXT_i),
    308                                                   MI->getDebugLoc());
    309 
    310   if (ResourceTracker->canReserveResources(PseudoMI)) {
    311     ResourceTracker->reserveResources(PseudoMI);
    312     MI->getParent()->getParent()->DeleteMachineInstr(PseudoMI);
    313     return true;
    314   } else {
    315     MI->getParent()->getParent()->DeleteMachineInstr(PseudoMI);
    316     return false;
    317   }
    318 }
    319 
    320 
    321 bool HexagonPacketizerList::IsCallDependent(MachineInstr* MI,
    322                                           SDep::Kind DepType,
    323                                           unsigned DepReg) {
    324 
    325   const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
    326   const HexagonRegisterInfo* QRI =
    327               (const HexagonRegisterInfo *) TM.getRegisterInfo();
    328 
    329   // Check for lr dependence
    330   if (DepReg == QRI->getRARegister()) {
    331     return true;
    332   }
    333 
    334   if (QII->isDeallocRet(MI)) {
    335     if (DepReg == QRI->getFrameRegister() ||
    336         DepReg == QRI->getStackRegister())
    337       return true;
    338   }
    339 
    340   // Check if this is a predicate dependence
    341   const TargetRegisterClass* RC = QRI->getMinimalPhysRegClass(DepReg);
    342   if (RC == &Hexagon::PredRegsRegClass) {
    343     return true;
    344   }
    345 
    346   //
    347   // Lastly check for an operand used in an indirect call
    348   // If we had an attribute for checking if an instruction is an indirect call,
    349   // then we could have avoided this relatively brittle implementation of
    350   // IsIndirectCall()
    351   //
    352   // Assumes that the first operand of the CALLr is the function address
    353   //
    354   if (IsIndirectCall(MI) && (DepType == SDep::Data)) {
    355     MachineOperand MO = MI->getOperand(0);
    356     if (MO.isReg() && MO.isUse() && (MO.getReg() == DepReg)) {
    357       return true;
    358     }
    359   }
    360 
    361   return false;
    362 }
    363 
    364 static bool IsRegDependence(const SDep::Kind DepType) {
    365   return (DepType == SDep::Data || DepType == SDep::Anti ||
    366           DepType == SDep::Output);
    367 }
    368 
    369 static bool IsDirectJump(MachineInstr* MI) {
    370   return (MI->getOpcode() == Hexagon::JMP);
    371 }
    372 
    373 static bool IsSchedBarrier(MachineInstr* MI) {
    374   switch (MI->getOpcode()) {
    375   case Hexagon::BARRIER:
    376     return true;
    377   }
    378   return false;
    379 }
    380 
    381 static bool IsControlFlow(MachineInstr* MI) {
    382   return (MI->getDesc().isTerminator() || MI->getDesc().isCall());
    383 }
    384 
    385 static bool IsLoopN(MachineInstr *MI) {
    386   return (MI->getOpcode() == Hexagon::LOOP0_i ||
    387           MI->getOpcode() == Hexagon::LOOP0_r);
    388 }
    389 
    390 /// DoesModifyCalleeSavedReg - Returns true if the instruction modifies a
    391 /// callee-saved register.
    392 static bool DoesModifyCalleeSavedReg(MachineInstr *MI,
    393                                      const TargetRegisterInfo *TRI) {
    394   for (const uint16_t *CSR = TRI->getCalleeSavedRegs(); *CSR; ++CSR) {
    395     unsigned CalleeSavedReg = *CSR;
    396     if (MI->modifiesRegister(CalleeSavedReg, TRI))
    397       return true;
    398   }
    399   return false;
    400 }
    401 
    402 // Returns true if an instruction can be promoted to .new predicate
    403 // or new-value store.
    404 bool HexagonPacketizerList::isNewifiable(MachineInstr* MI) {
    405   const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
    406   if ( isCondInst(MI) || QII->mayBeNewStore(MI))
    407     return true;
    408   else
    409     return false;
    410 }
    411 
    412 bool HexagonPacketizerList::isCondInst (MachineInstr* MI) {
    413   const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
    414   const MCInstrDesc& TID = MI->getDesc();
    415                                     // bug 5670: until that is fixed,
    416                                     // this portion is disabled.
    417   if (   TID.isConditionalBranch()  // && !IsRegisterJump(MI)) ||
    418       || QII->isConditionalTransfer(MI)
    419       || QII->isConditionalALU32(MI)
    420       || QII->isConditionalLoad(MI)
    421       || QII->isConditionalStore(MI)) {
    422     return true;
    423   }
    424   return false;
    425 }
    426 
    427 
    428 // Promote an instructiont to its .new form.
    429 // At this time, we have already made a call to CanPromoteToDotNew
    430 // and made sure that it can *indeed* be promoted.
    431 bool HexagonPacketizerList::PromoteToDotNew(MachineInstr* MI,
    432                         SDep::Kind DepType, MachineBasicBlock::iterator &MII,
    433                         const TargetRegisterClass* RC) {
    434 
    435   assert (DepType == SDep::Data);
    436   const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
    437 
    438   int NewOpcode;
    439   if (RC == &Hexagon::PredRegsRegClass)
    440     NewOpcode = QII->GetDotNewPredOp(MI, MBPI);
    441   else
    442     NewOpcode = QII->GetDotNewOp(MI);
    443   MI->setDesc(QII->get(NewOpcode));
    444 
    445   return true;
    446 }
    447 
    448 bool HexagonPacketizerList::DemoteToDotOld(MachineInstr* MI) {
    449   const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
    450   int NewOpcode = QII->GetDotOldOp(MI->getOpcode());
    451   MI->setDesc(QII->get(NewOpcode));
    452   return true;
    453 }
    454 
    455 enum PredicateKind {
    456   PK_False,
    457   PK_True,
    458   PK_Unknown
    459 };
    460 
    461 /// Returns true if an instruction is predicated on p0 and false if it's
    462 /// predicated on !p0.
    463 static PredicateKind getPredicateSense(MachineInstr* MI,
    464                                        const HexagonInstrInfo *QII) {
    465   if (!QII->isPredicated(MI))
    466     return PK_Unknown;
    467 
    468   if (QII->isPredicatedTrue(MI))
    469     return PK_True;
    470 
    471   return PK_False;
    472 }
    473 
    474 static MachineOperand& GetPostIncrementOperand(MachineInstr *MI,
    475                                                const HexagonInstrInfo *QII) {
    476   assert(QII->isPostIncrement(MI) && "Not a post increment operation.");
    477 #ifndef NDEBUG
    478   // Post Increment means duplicates. Use dense map to find duplicates in the
    479   // list. Caution: Densemap initializes with the minimum of 64 buckets,
    480   // whereas there are at most 5 operands in the post increment.
    481   DenseMap<unsigned,  unsigned> DefRegsSet;
    482   for(unsigned opNum = 0; opNum < MI->getNumOperands(); opNum++)
    483     if (MI->getOperand(opNum).isReg() &&
    484         MI->getOperand(opNum).isDef()) {
    485       DefRegsSet[MI->getOperand(opNum).getReg()] = 1;
    486     }
    487 
    488   for(unsigned opNum = 0; opNum < MI->getNumOperands(); opNum++)
    489     if (MI->getOperand(opNum).isReg() &&
    490         MI->getOperand(opNum).isUse()) {
    491       if (DefRegsSet[MI->getOperand(opNum).getReg()]) {
    492         return MI->getOperand(opNum);
    493       }
    494     }
    495 #else
    496   if (MI->getDesc().mayLoad()) {
    497     // The 2nd operand is always the post increment operand in load.
    498     assert(MI->getOperand(1).isReg() &&
    499                 "Post increment operand has be to a register.");
    500     return (MI->getOperand(1));
    501   }
    502   if (MI->getDesc().mayStore()) {
    503     // The 1st operand is always the post increment operand in store.
    504     assert(MI->getOperand(0).isReg() &&
    505                 "Post increment operand has be to a register.");
    506     return (MI->getOperand(0));
    507   }
    508 #endif
    509   // we should never come here.
    510   llvm_unreachable("mayLoad or mayStore not set for Post Increment operation");
    511 }
    512 
    513 // get the value being stored
    514 static MachineOperand& GetStoreValueOperand(MachineInstr *MI) {
    515   // value being stored is always the last operand.
    516   return (MI->getOperand(MI->getNumOperands()-1));
    517 }
    518 
    519 // can be new value store?
    520 // Following restrictions are to be respected in convert a store into
    521 // a new value store.
    522 // 1. If an instruction uses auto-increment, its address register cannot
    523 //    be a new-value register. Arch Spec 5.4.2.1
    524 // 2. If an instruction uses absolute-set addressing mode,
    525 //    its address register cannot be a new-value register.
    526 //    Arch Spec 5.4.2.1.TODO: This is not enabled as
    527 //    as absolute-set address mode patters are not implemented.
    528 // 3. If an instruction produces a 64-bit result, its registers cannot be used
    529 //    as new-value registers. Arch Spec 5.4.2.2.
    530 // 4. If the instruction that sets a new-value register is conditional, then
    531 //    the instruction that uses the new-value register must also be conditional,
    532 //    and both must always have their predicates evaluate identically.
    533 //    Arch Spec 5.4.2.3.
    534 // 5. There is an implied restriction of a packet can not have another store,
    535 //    if there is a  new value store in the packet. Corollary, if there is
    536 //    already a store in a packet, there can not be a new value store.
    537 //    Arch Spec: 3.4.4.2
    538 bool HexagonPacketizerList::CanPromoteToNewValueStore( MachineInstr *MI,
    539                 MachineInstr *PacketMI, unsigned DepReg,
    540                 std::map <MachineInstr*, SUnit*> MIToSUnit) {
    541   const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
    542   // Make sure we are looking at the store, that can be promoted.
    543   if (!QII->mayBeNewStore(MI))
    544     return false;
    545 
    546   // Make sure there is dependency and can be new value'ed
    547   if (GetStoreValueOperand(MI).isReg() &&
    548       GetStoreValueOperand(MI).getReg() != DepReg)
    549     return false;
    550 
    551   const HexagonRegisterInfo* QRI =
    552                             (const HexagonRegisterInfo *) TM.getRegisterInfo();
    553   const MCInstrDesc& MCID = PacketMI->getDesc();
    554   // first operand is always the result
    555 
    556   const TargetRegisterClass* PacketRC = QII->getRegClass(MCID, 0, QRI, MF);
    557 
    558   // if there is already an store in the packet, no can do new value store
    559   // Arch Spec 3.4.4.2.
    560   for (std::vector<MachineInstr*>::iterator VI = CurrentPacketMIs.begin(),
    561          VE = CurrentPacketMIs.end();
    562        (VI != VE); ++VI) {
    563     SUnit* PacketSU = MIToSUnit[*VI];
    564     if (PacketSU->getInstr()->getDesc().mayStore() ||
    565         // if we have mayStore = 1 set on ALLOCFRAME and DEALLOCFRAME,
    566         // then we don't need this
    567         PacketSU->getInstr()->getOpcode() == Hexagon::ALLOCFRAME ||
    568         PacketSU->getInstr()->getOpcode() == Hexagon::DEALLOCFRAME)
    569       return false;
    570   }
    571 
    572   if (PacketRC == &Hexagon::DoubleRegsRegClass) {
    573     // new value store constraint: double regs can not feed into new value store
    574     // arch spec section: 5.4.2.2
    575     return false;
    576   }
    577 
    578   // Make sure it's NOT the post increment register that we are going to
    579   // new value.
    580   if (QII->isPostIncrement(MI) &&
    581       MI->getDesc().mayStore() &&
    582       GetPostIncrementOperand(MI, QII).getReg() == DepReg) {
    583     return false;
    584   }
    585 
    586   if (QII->isPostIncrement(PacketMI) &&
    587       PacketMI->getDesc().mayLoad() &&
    588       GetPostIncrementOperand(PacketMI, QII).getReg() == DepReg) {
    589     // if source is post_inc, or absolute-set addressing,
    590     // it can not feed into new value store
    591     //  r3 = memw(r2++#4)
    592     //  memw(r30 + #-1404) = r2.new -> can not be new value store
    593     // arch spec section: 5.4.2.1
    594     return false;
    595   }
    596 
    597   // If the source that feeds the store is predicated, new value store must
    598   // also be predicated.
    599   if (QII->isPredicated(PacketMI)) {
    600     if (!QII->isPredicated(MI))
    601       return false;
    602 
    603     // Check to make sure that they both will have their predicates
    604     // evaluate identically
    605     unsigned predRegNumSrc = 0;
    606     unsigned predRegNumDst = 0;
    607     const TargetRegisterClass* predRegClass = NULL;
    608 
    609     // Get predicate register used in the source instruction
    610     for(unsigned opNum = 0; opNum < PacketMI->getNumOperands(); opNum++) {
    611       if ( PacketMI->getOperand(opNum).isReg())
    612       predRegNumSrc = PacketMI->getOperand(opNum).getReg();
    613       predRegClass = QRI->getMinimalPhysRegClass(predRegNumSrc);
    614       if (predRegClass == &Hexagon::PredRegsRegClass) {
    615         break;
    616       }
    617     }
    618     assert ((predRegClass == &Hexagon::PredRegsRegClass ) &&
    619         ("predicate register not found in a predicated PacketMI instruction"));
    620 
    621     // Get predicate register used in new-value store instruction
    622     for(unsigned opNum = 0; opNum < MI->getNumOperands(); opNum++) {
    623       if ( MI->getOperand(opNum).isReg())
    624       predRegNumDst = MI->getOperand(opNum).getReg();
    625       predRegClass = QRI->getMinimalPhysRegClass(predRegNumDst);
    626       if (predRegClass == &Hexagon::PredRegsRegClass) {
    627         break;
    628       }
    629     }
    630     assert ((predRegClass == &Hexagon::PredRegsRegClass ) &&
    631             ("predicate register not found in a predicated MI instruction"));
    632 
    633     // New-value register producer and user (store) need to satisfy these
    634     // constraints:
    635     // 1) Both instructions should be predicated on the same register.
    636     // 2) If producer of the new-value register is .new predicated then store
    637     // should also be .new predicated and if producer is not .new predicated
    638     // then store should not be .new predicated.
    639     // 3) Both new-value register producer and user should have same predicate
    640     // sense, i.e, either both should be negated or both should be none negated.
    641 
    642     if (( predRegNumDst != predRegNumSrc) ||
    643           QII->isDotNewInst(PacketMI) != QII->isDotNewInst(MI)  ||
    644           getPredicateSense(MI, QII) != getPredicateSense(PacketMI, QII)) {
    645       return false;
    646     }
    647   }
    648 
    649   // Make sure that other than the new-value register no other store instruction
    650   // register has been modified in the same packet. Predicate registers can be
    651   // modified by they should not be modified between the producer and the store
    652   // instruction as it will make them both conditional on different values.
    653   // We already know this to be true for all the instructions before and
    654   // including PacketMI. Howerver, we need to perform the check for the
    655   // remaining instructions in the packet.
    656 
    657   std::vector<MachineInstr*>::iterator VI;
    658   std::vector<MachineInstr*>::iterator VE;
    659   unsigned StartCheck = 0;
    660 
    661   for (VI=CurrentPacketMIs.begin(), VE = CurrentPacketMIs.end();
    662       (VI != VE); ++VI) {
    663     SUnit* TempSU = MIToSUnit[*VI];
    664     MachineInstr* TempMI = TempSU->getInstr();
    665 
    666     // Following condition is true for all the instructions until PacketMI is
    667     // reached (StartCheck is set to 0 before the for loop).
    668     // StartCheck flag is 1 for all the instructions after PacketMI.
    669     if (TempMI != PacketMI && !StartCheck) // start processing only after
    670       continue;                            // encountering PacketMI
    671 
    672     StartCheck = 1;
    673     if (TempMI == PacketMI) // We don't want to check PacketMI for dependence
    674       continue;
    675 
    676     for(unsigned opNum = 0; opNum < MI->getNumOperands(); opNum++) {
    677       if (MI->getOperand(opNum).isReg() &&
    678           TempSU->getInstr()->modifiesRegister(MI->getOperand(opNum).getReg(),
    679                                                QRI))
    680         return false;
    681     }
    682   }
    683 
    684   // Make sure that for non POST_INC stores:
    685   // 1. The only use of reg is DepReg and no other registers.
    686   //    This handles V4 base+index registers.
    687   //    The following store can not be dot new.
    688   //    Eg.   r0 = add(r0, #3)a
    689   //          memw(r1+r0<<#2) = r0
    690   if (!QII->isPostIncrement(MI) &&
    691       GetStoreValueOperand(MI).isReg() &&
    692       GetStoreValueOperand(MI).getReg() == DepReg) {
    693     for(unsigned opNum = 0; opNum < MI->getNumOperands()-1; opNum++) {
    694       if (MI->getOperand(opNum).isReg() &&
    695           MI->getOperand(opNum).getReg() == DepReg) {
    696         return false;
    697       }
    698     }
    699     // 2. If data definition is because of implicit definition of the register,
    700     //    do not newify the store. Eg.
    701     //    %R9<def> = ZXTH %R12, %D6<imp-use>, %R12<imp-def>
    702     //    STrih_indexed %R8, 2, %R12<kill>; mem:ST2[%scevgep343]
    703     for(unsigned opNum = 0; opNum < PacketMI->getNumOperands(); opNum++) {
    704       if (PacketMI->getOperand(opNum).isReg() &&
    705           PacketMI->getOperand(opNum).getReg() == DepReg &&
    706           PacketMI->getOperand(opNum).isDef() &&
    707           PacketMI->getOperand(opNum).isImplicit()) {
    708         return false;
    709       }
    710     }
    711   }
    712 
    713   // Can be dot new store.
    714   return true;
    715 }
    716 
    717 // can this MI to promoted to either
    718 // new value store or new value jump
    719 bool HexagonPacketizerList::CanPromoteToNewValue( MachineInstr *MI,
    720                 SUnit *PacketSU, unsigned DepReg,
    721                 std::map <MachineInstr*, SUnit*> MIToSUnit,
    722                 MachineBasicBlock::iterator &MII)
    723 {
    724 
    725   const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
    726   const HexagonRegisterInfo* QRI =
    727                             (const HexagonRegisterInfo *) TM.getRegisterInfo();
    728   if (!QRI->Subtarget.hasV4TOps() ||
    729       !QII->mayBeNewStore(MI))
    730     return false;
    731 
    732   MachineInstr *PacketMI = PacketSU->getInstr();
    733 
    734   // Check to see the store can be new value'ed.
    735   if (CanPromoteToNewValueStore(MI, PacketMI, DepReg, MIToSUnit))
    736     return true;
    737 
    738   // Check to see the compare/jump can be new value'ed.
    739   // This is done as a pass on its own. Don't need to check it here.
    740   return false;
    741 }
    742 
    743 // Check to see if an instruction can be dot new
    744 // There are three kinds.
    745 // 1. dot new on predicate - V2/V3/V4
    746 // 2. dot new on stores NV/ST - V4
    747 // 3. dot new on jump NV/J - V4 -- This is generated in a pass.
    748 bool HexagonPacketizerList::CanPromoteToDotNew( MachineInstr *MI,
    749                               SUnit *PacketSU, unsigned DepReg,
    750                               std::map <MachineInstr*, SUnit*> MIToSUnit,
    751                               MachineBasicBlock::iterator &MII,
    752                               const TargetRegisterClass* RC )
    753 {
    754   const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
    755   // Already a dot new instruction.
    756   if (QII->isDotNewInst(MI) && !QII->mayBeNewStore(MI))
    757     return false;
    758 
    759   if (!isNewifiable(MI))
    760     return false;
    761 
    762   // predicate .new
    763   if (RC == &Hexagon::PredRegsRegClass && isCondInst(MI))
    764       return true;
    765   else if (RC != &Hexagon::PredRegsRegClass &&
    766       !QII->mayBeNewStore(MI)) // MI is not a new-value store
    767     return false;
    768   else {
    769     // Create a dot new machine instruction to see if resources can be
    770     // allocated. If not, bail out now.
    771     int NewOpcode = QII->GetDotNewOp(MI);
    772     const MCInstrDesc &desc = QII->get(NewOpcode);
    773     DebugLoc dl;
    774     MachineInstr *NewMI =
    775                     MI->getParent()->getParent()->CreateMachineInstr(desc, dl);
    776     bool ResourcesAvailable = ResourceTracker->canReserveResources(NewMI);
    777     MI->getParent()->getParent()->DeleteMachineInstr(NewMI);
    778 
    779     if (!ResourcesAvailable)
    780       return false;
    781 
    782     // new value store only
    783     // new new value jump generated as a passes
    784     if (!CanPromoteToNewValue(MI, PacketSU, DepReg, MIToSUnit, MII)) {
    785       return false;
    786     }
    787   }
    788   return true;
    789 }
    790 
    791 // Go through the packet instructions and search for anti dependency
    792 // between them and DepReg from MI
    793 // Consider this case:
    794 // Trying to add
    795 // a) %R1<def> = TFRI_cdNotPt %P3, 2
    796 // to this packet:
    797 // {
    798 //   b) %P0<def> = OR_pp %P3<kill>, %P0<kill>
    799 //   c) %P3<def> = TFR_PdRs %R23
    800 //   d) %R1<def> = TFRI_cdnPt %P3, 4
    801 //  }
    802 // The P3 from a) and d) will be complements after
    803 // a)'s P3 is converted to .new form
    804 // Anti Dep between c) and b) is irrelevant for this case
    805 bool HexagonPacketizerList::RestrictingDepExistInPacket (MachineInstr* MI,
    806       unsigned DepReg,
    807       std::map <MachineInstr*, SUnit*> MIToSUnit) {
    808 
    809   const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
    810   SUnit* PacketSUDep = MIToSUnit[MI];
    811 
    812   for (std::vector<MachineInstr*>::iterator VIN = CurrentPacketMIs.begin(),
    813        VEN = CurrentPacketMIs.end(); (VIN != VEN); ++VIN) {
    814 
    815     // We only care for dependencies to predicated instructions
    816     if(!QII->isPredicated(*VIN)) continue;
    817 
    818     // Scheduling Unit for current insn in the packet
    819     SUnit* PacketSU = MIToSUnit[*VIN];
    820 
    821     // Look at dependencies between current members of the packet
    822     // and predicate defining instruction MI.
    823     // Make sure that dependency is on the exact register
    824     // we care about.
    825     if (PacketSU->isSucc(PacketSUDep)) {
    826       for (unsigned i = 0; i < PacketSU->Succs.size(); ++i) {
    827         if ((PacketSU->Succs[i].getSUnit() == PacketSUDep) &&
    828             (PacketSU->Succs[i].getKind() == SDep::Anti) &&
    829             (PacketSU->Succs[i].getReg() == DepReg)) {
    830           return true;
    831         }
    832       }
    833     }
    834   }
    835 
    836   return false;
    837 }
    838 
    839 
    840 /// Gets the predicate register of a predicated instruction.
    841 static unsigned getPredicatedRegister(MachineInstr *MI,
    842                                       const HexagonInstrInfo *QII) {
    843   /// We use the following rule: The first predicate register that is a use is
    844   /// the predicate register of a predicated instruction.
    845 
    846   assert(QII->isPredicated(MI) && "Must be predicated instruction");
    847 
    848   for (MachineInstr::mop_iterator OI = MI->operands_begin(),
    849        OE = MI->operands_end(); OI != OE; ++OI) {
    850     MachineOperand &Op = *OI;
    851     if (Op.isReg() && Op.getReg() && Op.isUse() &&
    852         Hexagon::PredRegsRegClass.contains(Op.getReg()))
    853       return Op.getReg();
    854   }
    855 
    856   llvm_unreachable("Unknown instruction operand layout");
    857 
    858   return 0;
    859 }
    860 
    861 // Given two predicated instructions, this function detects whether
    862 // the predicates are complements
    863 bool HexagonPacketizerList::ArePredicatesComplements (MachineInstr* MI1,
    864      MachineInstr* MI2, std::map <MachineInstr*, SUnit*> MIToSUnit) {
    865 
    866   const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
    867 
    868   // If we don't know the predicate sense of the instructions bail out early, we
    869   // need it later.
    870   if (getPredicateSense(MI1, QII) == PK_Unknown ||
    871       getPredicateSense(MI2, QII) == PK_Unknown)
    872     return false;
    873 
    874   // Scheduling unit for candidate
    875   SUnit* SU = MIToSUnit[MI1];
    876 
    877   // One corner case deals with the following scenario:
    878   // Trying to add
    879   // a) %R24<def> = TFR_cPt %P0, %R25
    880   // to this packet:
    881   //
    882   // {
    883   //   b) %R25<def> = TFR_cNotPt %P0, %R24
    884   //   c) %P0<def> = CMPEQri %R26, 1
    885   // }
    886   //
    887   // On general check a) and b) are complements, but
    888   // presence of c) will convert a) to .new form, and
    889   // then it is not a complement
    890   // We attempt to detect it by analyzing  existing
    891   // dependencies in the packet
    892 
    893   // Analyze relationships between all existing members of the packet.
    894   // Look for Anti dependecy on the same predicate reg
    895   // as used in the candidate
    896   for (std::vector<MachineInstr*>::iterator VIN = CurrentPacketMIs.begin(),
    897        VEN = CurrentPacketMIs.end(); (VIN != VEN); ++VIN) {
    898 
    899     // Scheduling Unit for current insn in the packet
    900     SUnit* PacketSU = MIToSUnit[*VIN];
    901 
    902     // If this instruction in the packet is succeeded by the candidate...
    903     if (PacketSU->isSucc(SU)) {
    904       for (unsigned i = 0; i < PacketSU->Succs.size(); ++i) {
    905         // The corner case exist when there is true data
    906         // dependency between candidate and one of current
    907         // packet members, this dep is on predicate reg, and
    908         // there already exist anti dep on the same pred in
    909         // the packet.
    910         if (PacketSU->Succs[i].getSUnit() == SU &&
    911             PacketSU->Succs[i].getKind() == SDep::Data &&
    912             Hexagon::PredRegsRegClass.contains(
    913               PacketSU->Succs[i].getReg()) &&
    914             // Here I know that *VIN is predicate setting instruction
    915             // with true data dep to candidate on the register
    916             // we care about - c) in the above example.
    917             // Now I need to see if there is an anti dependency
    918             // from c) to any other instruction in the
    919             // same packet on the pred reg of interest
    920             RestrictingDepExistInPacket(*VIN,PacketSU->Succs[i].getReg(),
    921                                         MIToSUnit)) {
    922            return false;
    923         }
    924       }
    925     }
    926   }
    927 
    928   // If the above case does not apply, check regular
    929   // complement condition.
    930   // Check that the predicate register is the same and
    931   // that the predicate sense is different
    932   // We also need to differentiate .old vs. .new:
    933   // !p0 is not complimentary to p0.new
    934   unsigned PReg1 = getPredicatedRegister(MI1, QII);
    935   unsigned PReg2 = getPredicatedRegister(MI2, QII);
    936   return ((PReg1 == PReg2) &&
    937           Hexagon::PredRegsRegClass.contains(PReg1) &&
    938           Hexagon::PredRegsRegClass.contains(PReg2) &&
    939           (getPredicateSense(MI1, QII) != getPredicateSense(MI2, QII)) &&
    940           (QII->isDotNewInst(MI1) == QII->isDotNewInst(MI2)));
    941 }
    942 
    943 // initPacketizerState - Initialize packetizer flags
    944 void HexagonPacketizerList::initPacketizerState() {
    945 
    946   Dependence = false;
    947   PromotedToDotNew = false;
    948   GlueToNewValueJump = false;
    949   GlueAllocframeStore = false;
    950   FoundSequentialDependence = false;
    951 
    952   return;
    953 }
    954 
    955 // ignorePseudoInstruction - Ignore bundling of pseudo instructions.
    956 bool HexagonPacketizerList::ignorePseudoInstruction(MachineInstr *MI,
    957                                                     MachineBasicBlock *MBB) {
    958   if (MI->isDebugValue())
    959     return true;
    960 
    961   // We must print out inline assembly
    962   if (MI->isInlineAsm())
    963     return false;
    964 
    965   // We check if MI has any functional units mapped to it.
    966   // If it doesn't, we ignore the instruction.
    967   const MCInstrDesc& TID = MI->getDesc();
    968   unsigned SchedClass = TID.getSchedClass();
    969   const InstrStage* IS =
    970                     ResourceTracker->getInstrItins()->beginStage(SchedClass);
    971   unsigned FuncUnits = IS->getUnits();
    972   return !FuncUnits;
    973 }
    974 
    975 // isSoloInstruction: - Returns true for instructions that must be
    976 // scheduled in their own packet.
    977 bool HexagonPacketizerList::isSoloInstruction(MachineInstr *MI) {
    978 
    979   if (MI->isInlineAsm())
    980     return true;
    981 
    982   if (MI->isEHLabel())
    983     return true;
    984 
    985   // From Hexagon V4 Programmer's Reference Manual 3.4.4 Grouping constraints:
    986   // trap, pause, barrier, icinva, isync, and syncht are solo instructions.
    987   // They must not be grouped with other instructions in a packet.
    988   if (IsSchedBarrier(MI))
    989     return true;
    990 
    991   return false;
    992 }
    993 
    994 // isLegalToPacketizeTogether:
    995 // SUI is the current instruction that is out side of the current packet.
    996 // SUJ is the current instruction inside the current packet against which that
    997 // SUI will be packetized.
    998 bool HexagonPacketizerList::isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) {
    999   MachineInstr *I = SUI->getInstr();
   1000   MachineInstr *J = SUJ->getInstr();
   1001   assert(I && J && "Unable to packetize null instruction!");
   1002 
   1003   const MCInstrDesc &MCIDI = I->getDesc();
   1004   const MCInstrDesc &MCIDJ = J->getDesc();
   1005 
   1006   MachineBasicBlock::iterator II = I;
   1007 
   1008   const unsigned FrameSize = MF.getFrameInfo()->getStackSize();
   1009   const HexagonRegisterInfo* QRI =
   1010                       (const HexagonRegisterInfo *) TM.getRegisterInfo();
   1011   const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
   1012 
   1013   // Inline asm cannot go in the packet.
   1014   if (I->getOpcode() == Hexagon::INLINEASM)
   1015     llvm_unreachable("Should not meet inline asm here!");
   1016 
   1017   if (isSoloInstruction(I))
   1018     llvm_unreachable("Should not meet solo instr here!");
   1019 
   1020   // A save callee-save register function call can only be in a packet
   1021   // with instructions that don't write to the callee-save registers.
   1022   if ((QII->isSaveCalleeSavedRegsCall(I) &&
   1023        DoesModifyCalleeSavedReg(J, QRI)) ||
   1024       (QII->isSaveCalleeSavedRegsCall(J) &&
   1025        DoesModifyCalleeSavedReg(I, QRI))) {
   1026     Dependence = true;
   1027     return false;
   1028   }
   1029 
   1030   // Two control flow instructions cannot go in the same packet.
   1031   if (IsControlFlow(I) && IsControlFlow(J)) {
   1032     Dependence = true;
   1033     return false;
   1034   }
   1035 
   1036   // A LoopN instruction cannot appear in the same packet as a jump or call.
   1037   if (IsLoopN(I) &&
   1038      (IsDirectJump(J) || MCIDJ.isCall() || QII->isDeallocRet(J))) {
   1039     Dependence = true;
   1040     return false;
   1041   }
   1042   if (IsLoopN(J) &&
   1043      (IsDirectJump(I) || MCIDI.isCall() || QII->isDeallocRet(I))) {
   1044     Dependence = true;
   1045     return false;
   1046   }
   1047 
   1048   // dealloc_return cannot appear in the same packet as a conditional or
   1049   // unconditional jump.
   1050   if (QII->isDeallocRet(I) &&
   1051      (MCIDJ.isBranch() || MCIDJ.isCall() || MCIDJ.isBarrier())) {
   1052     Dependence = true;
   1053     return false;
   1054   }
   1055 
   1056 
   1057   // V4 allows dual store. But does not allow second store, if the
   1058   // first store is not in SLOT0. New value store, new value jump,
   1059   // dealloc_return and memop always take SLOT0.
   1060   // Arch spec 3.4.4.2
   1061   if (QRI->Subtarget.hasV4TOps()) {
   1062     if (MCIDI.mayStore() && MCIDJ.mayStore() &&
   1063        (QII->isNewValueInst(J) || QII->isMemOp(J) || QII->isMemOp(I))) {
   1064       Dependence = true;
   1065       return false;
   1066     }
   1067 
   1068     if ((QII->isMemOp(J) && MCIDI.mayStore())
   1069         || (MCIDJ.mayStore() && QII->isMemOp(I))
   1070         || (QII->isMemOp(J) && QII->isMemOp(I))) {
   1071       Dependence = true;
   1072       return false;
   1073     }
   1074 
   1075     //if dealloc_return
   1076     if (MCIDJ.mayStore() && QII->isDeallocRet(I)) {
   1077       Dependence = true;
   1078       return false;
   1079     }
   1080 
   1081     // If an instruction feeds new value jump, glue it.
   1082     MachineBasicBlock::iterator NextMII = I;
   1083     ++NextMII;
   1084     if (NextMII != I->getParent()->end() && QII->isNewValueJump(NextMII)) {
   1085       MachineInstr *NextMI = NextMII;
   1086 
   1087       bool secondRegMatch = false;
   1088       bool maintainNewValueJump = false;
   1089 
   1090       if (NextMI->getOperand(1).isReg() &&
   1091           I->getOperand(0).getReg() == NextMI->getOperand(1).getReg()) {
   1092         secondRegMatch = true;
   1093         maintainNewValueJump = true;
   1094       }
   1095 
   1096       if (!secondRegMatch &&
   1097            I->getOperand(0).getReg() == NextMI->getOperand(0).getReg()) {
   1098         maintainNewValueJump = true;
   1099       }
   1100 
   1101       for (std::vector<MachineInstr*>::iterator
   1102             VI = CurrentPacketMIs.begin(),
   1103              VE = CurrentPacketMIs.end();
   1104            (VI != VE && maintainNewValueJump); ++VI) {
   1105         SUnit* PacketSU = MIToSUnit[*VI];
   1106 
   1107         // NVJ can not be part of the dual jump - Arch Spec: section 7.8
   1108         if (PacketSU->getInstr()->getDesc().isCall()) {
   1109           Dependence = true;
   1110           break;
   1111         }
   1112         // Validate
   1113         // 1. Packet does not have a store in it.
   1114         // 2. If the first operand of the nvj is newified, and the second
   1115         //    operand is also a reg, it (second reg) is not defined in
   1116         //    the same packet.
   1117         // 3. If the second operand of the nvj is newified, (which means
   1118         //    first operand is also a reg), first reg is not defined in
   1119         //    the same packet.
   1120         if (PacketSU->getInstr()->getDesc().mayStore()               ||
   1121             PacketSU->getInstr()->getOpcode() == Hexagon::ALLOCFRAME ||
   1122             // Check #2.
   1123             (!secondRegMatch && NextMI->getOperand(1).isReg() &&
   1124              PacketSU->getInstr()->modifiesRegister(
   1125                                NextMI->getOperand(1).getReg(), QRI)) ||
   1126             // Check #3.
   1127             (secondRegMatch &&
   1128              PacketSU->getInstr()->modifiesRegister(
   1129                                NextMI->getOperand(0).getReg(), QRI))) {
   1130           Dependence = true;
   1131           break;
   1132         }
   1133       }
   1134       if (!Dependence)
   1135         GlueToNewValueJump = true;
   1136       else
   1137         return false;
   1138     }
   1139   }
   1140 
   1141   if (SUJ->isSucc(SUI)) {
   1142     for (unsigned i = 0;
   1143          (i < SUJ->Succs.size()) && !FoundSequentialDependence;
   1144          ++i) {
   1145 
   1146       if (SUJ->Succs[i].getSUnit() != SUI) {
   1147         continue;
   1148       }
   1149 
   1150       SDep::Kind DepType = SUJ->Succs[i].getKind();
   1151 
   1152       // For direct calls:
   1153       // Ignore register dependences for call instructions for
   1154       // packetization purposes except for those due to r31 and
   1155       // predicate registers.
   1156       //
   1157       // For indirect calls:
   1158       // Same as direct calls + check for true dependences to the register
   1159       // used in the indirect call.
   1160       //
   1161       // We completely ignore Order dependences for call instructions
   1162       //
   1163       // For returns:
   1164       // Ignore register dependences for return instructions like jumpr,
   1165       // dealloc return unless we have dependencies on the explicit uses
   1166       // of the registers used by jumpr (like r31) or dealloc return
   1167       // (like r29 or r30).
   1168       //
   1169       // TODO: Currently, jumpr is handling only return of r31. So, the
   1170       // following logic (specificaly IsCallDependent) is working fine.
   1171       // We need to enable jumpr for register other than r31 and then,
   1172       // we need to rework the last part, where it handles indirect call
   1173       // of that (IsCallDependent) function. Bug 6216 is opened for this.
   1174       //
   1175       unsigned DepReg = 0;
   1176       const TargetRegisterClass* RC = NULL;
   1177       if (DepType == SDep::Data) {
   1178         DepReg = SUJ->Succs[i].getReg();
   1179         RC = QRI->getMinimalPhysRegClass(DepReg);
   1180       }
   1181       if ((MCIDI.isCall() || MCIDI.isReturn()) &&
   1182           (!IsRegDependence(DepType) ||
   1183             !IsCallDependent(I, DepType, SUJ->Succs[i].getReg()))) {
   1184         /* do nothing */
   1185       }
   1186 
   1187       // For instructions that can be promoted to dot-new, try to promote.
   1188       else if ((DepType == SDep::Data) &&
   1189                CanPromoteToDotNew(I, SUJ, DepReg, MIToSUnit, II, RC) &&
   1190                PromoteToDotNew(I, DepType, II, RC)) {
   1191         PromotedToDotNew = true;
   1192         /* do nothing */
   1193       }
   1194 
   1195       else if ((DepType == SDep::Data) &&
   1196                (QII->isNewValueJump(I))) {
   1197         /* do nothing */
   1198       }
   1199 
   1200       // For predicated instructions, if the predicates are complements
   1201       // then there can be no dependence.
   1202       else if (QII->isPredicated(I) &&
   1203                QII->isPredicated(J) &&
   1204           ArePredicatesComplements(I, J, MIToSUnit)) {
   1205         /* do nothing */
   1206 
   1207       }
   1208       else if (IsDirectJump(I) &&
   1209                !MCIDJ.isBranch() &&
   1210                !MCIDJ.isCall() &&
   1211                (DepType == SDep::Order)) {
   1212         // Ignore Order dependences between unconditional direct branches
   1213         // and non-control-flow instructions
   1214         /* do nothing */
   1215       }
   1216       else if (MCIDI.isConditionalBranch() && (DepType != SDep::Data) &&
   1217                (DepType != SDep::Output)) {
   1218         // Ignore all dependences for jumps except for true and output
   1219         // dependences
   1220         /* do nothing */
   1221       }
   1222 
   1223       // Ignore output dependences due to superregs. We can
   1224       // write to two different subregisters of R1:0 for instance
   1225       // in the same cycle
   1226       //
   1227 
   1228       //
   1229       // Let the
   1230       // If neither I nor J defines DepReg, then this is a
   1231       // superfluous output dependence. The dependence must be of the
   1232       // form:
   1233       //  R0 = ...
   1234       //  R1 = ...
   1235       // and there is an output dependence between the two instructions
   1236       // with
   1237       // DepReg = D0
   1238       // We want to ignore these dependences.
   1239       // Ideally, the dependence constructor should annotate such
   1240       // dependences. We can then avoid this relatively expensive check.
   1241       //
   1242       else if (DepType == SDep::Output) {
   1243         // DepReg is the register that's responsible for the dependence.
   1244         unsigned DepReg = SUJ->Succs[i].getReg();
   1245 
   1246         // Check if I and J really defines DepReg.
   1247         if (I->definesRegister(DepReg) ||
   1248             J->definesRegister(DepReg)) {
   1249           FoundSequentialDependence = true;
   1250           break;
   1251         }
   1252       }
   1253 
   1254       // We ignore Order dependences for
   1255       // 1. Two loads unless they are volatile.
   1256       // 2. Two stores in V4 unless they are volatile.
   1257       else if ((DepType == SDep::Order) &&
   1258                !I->hasOrderedMemoryRef() &&
   1259                !J->hasOrderedMemoryRef()) {
   1260         if (QRI->Subtarget.hasV4TOps() &&
   1261             // hexagonv4 allows dual store.
   1262             MCIDI.mayStore() && MCIDJ.mayStore()) {
   1263           /* do nothing */
   1264         }
   1265         // store followed by store-- not OK on V2
   1266         // store followed by load -- not OK on all (OK if addresses
   1267         // are not aliased)
   1268         // load followed by store -- OK on all
   1269         // load followed by load  -- OK on all
   1270         else if ( !MCIDJ.mayStore()) {
   1271           /* do nothing */
   1272         }
   1273         else {
   1274           FoundSequentialDependence = true;
   1275           break;
   1276         }
   1277       }
   1278 
   1279       // For V4, special case ALLOCFRAME. Even though there is dependency
   1280       // between ALLOCAFRAME and subsequent store, allow it to be
   1281       // packetized in a same packet. This implies that the store is using
   1282       // caller's SP. Hense, offset needs to be updated accordingly.
   1283       else if (DepType == SDep::Data
   1284                && QRI->Subtarget.hasV4TOps()
   1285                && J->getOpcode() == Hexagon::ALLOCFRAME
   1286                && (I->getOpcode() == Hexagon::STrid
   1287                    || I->getOpcode() == Hexagon::STriw
   1288                    || I->getOpcode() == Hexagon::STrib)
   1289                && I->getOperand(0).getReg() == QRI->getStackRegister()
   1290                && QII->isValidOffset(I->getOpcode(),
   1291                                      I->getOperand(1).getImm() -
   1292                                      (FrameSize + HEXAGON_LRFP_SIZE)))
   1293       {
   1294         GlueAllocframeStore = true;
   1295         // Since this store is to be glued with allocframe in the same
   1296         // packet, it will use SP of the previous stack frame, i.e
   1297         // caller's SP. Therefore, we need to recalculate offset according
   1298         // to this change.
   1299         I->getOperand(1).setImm(I->getOperand(1).getImm() -
   1300                                         (FrameSize + HEXAGON_LRFP_SIZE));
   1301       }
   1302 
   1303       //
   1304       // Skip over anti-dependences. Two instructions that are
   1305       // anti-dependent can share a packet
   1306       //
   1307       else if (DepType != SDep::Anti) {
   1308         FoundSequentialDependence = true;
   1309         break;
   1310       }
   1311     }
   1312 
   1313     if (FoundSequentialDependence) {
   1314       Dependence = true;
   1315       return false;
   1316     }
   1317   }
   1318 
   1319   return true;
   1320 }
   1321 
   1322 // isLegalToPruneDependencies
   1323 bool HexagonPacketizerList::isLegalToPruneDependencies(SUnit *SUI, SUnit *SUJ) {
   1324   MachineInstr *I = SUI->getInstr();
   1325   assert(I && SUJ->getInstr() && "Unable to packetize null instruction!");
   1326 
   1327   const unsigned FrameSize = MF.getFrameInfo()->getStackSize();
   1328 
   1329   if (Dependence) {
   1330 
   1331     // Check if the instruction was promoted to a dot-new. If so, demote it
   1332     // back into a dot-old.
   1333     if (PromotedToDotNew) {
   1334       DemoteToDotOld(I);
   1335     }
   1336 
   1337     // Check if the instruction (must be a store) was glued with an Allocframe
   1338     // instruction. If so, restore its offset to its original value, i.e. use
   1339     // curent SP instead of caller's SP.
   1340     if (GlueAllocframeStore) {
   1341       I->getOperand(1).setImm(I->getOperand(1).getImm() +
   1342                                              FrameSize + HEXAGON_LRFP_SIZE);
   1343     }
   1344 
   1345     return false;
   1346   }
   1347   return true;
   1348 }
   1349 
   1350 MachineBasicBlock::iterator
   1351 HexagonPacketizerList::addToPacket(MachineInstr *MI) {
   1352 
   1353     MachineBasicBlock::iterator MII = MI;
   1354     MachineBasicBlock *MBB = MI->getParent();
   1355 
   1356     const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
   1357 
   1358     if (GlueToNewValueJump) {
   1359 
   1360       ++MII;
   1361       MachineInstr *nvjMI = MII;
   1362       assert(ResourceTracker->canReserveResources(MI));
   1363       ResourceTracker->reserveResources(MI);
   1364       if ((QII->isExtended(MI) || QII->isConstExtended(MI)) &&
   1365           !tryAllocateResourcesForConstExt(MI)) {
   1366         endPacket(MBB, MI);
   1367         ResourceTracker->reserveResources(MI);
   1368         assert(canReserveResourcesForConstExt(MI) &&
   1369                "Ensure that there is a slot");
   1370         reserveResourcesForConstExt(MI);
   1371         // Reserve resources for new value jump constant extender.
   1372         assert(canReserveResourcesForConstExt(MI) &&
   1373                "Ensure that there is a slot");
   1374         reserveResourcesForConstExt(nvjMI);
   1375         assert(ResourceTracker->canReserveResources(nvjMI) &&
   1376                "Ensure that there is a slot");
   1377 
   1378       } else if (   // Extended instruction takes two slots in the packet.
   1379         // Try reserve and allocate 4-byte in the current packet first.
   1380         (QII->isExtended(nvjMI)
   1381             && (!tryAllocateResourcesForConstExt(nvjMI)
   1382                 || !ResourceTracker->canReserveResources(nvjMI)))
   1383         || // For non-extended instruction, no need to allocate extra 4 bytes.
   1384         (!QII->isExtended(nvjMI) &&
   1385               !ResourceTracker->canReserveResources(nvjMI)))
   1386       {
   1387         endPacket(MBB, MI);
   1388         // A new and empty packet starts.
   1389         // We are sure that the resources requirements can be satisfied.
   1390         // Therefore, do not need to call "canReserveResources" anymore.
   1391         ResourceTracker->reserveResources(MI);
   1392         if (QII->isExtended(nvjMI))
   1393           reserveResourcesForConstExt(nvjMI);
   1394       }
   1395       // Here, we are sure that "reserveResources" would succeed.
   1396       ResourceTracker->reserveResources(nvjMI);
   1397       CurrentPacketMIs.push_back(MI);
   1398       CurrentPacketMIs.push_back(nvjMI);
   1399     } else {
   1400       if (   (QII->isExtended(MI) || QII->isConstExtended(MI))
   1401           && (   !tryAllocateResourcesForConstExt(MI)
   1402               || !ResourceTracker->canReserveResources(MI)))
   1403       {
   1404         endPacket(MBB, MI);
   1405         // Check if the instruction was promoted to a dot-new. If so, demote it
   1406         // back into a dot-old
   1407         if (PromotedToDotNew) {
   1408           DemoteToDotOld(MI);
   1409         }
   1410         reserveResourcesForConstExt(MI);
   1411       }
   1412       // In case that "MI" is not an extended insn,
   1413       // the resource availability has already been checked.
   1414       ResourceTracker->reserveResources(MI);
   1415       CurrentPacketMIs.push_back(MI);
   1416     }
   1417     return MII;
   1418 }
   1419 
   1420 //===----------------------------------------------------------------------===//
   1421 //                         Public Constructor Functions
   1422 //===----------------------------------------------------------------------===//
   1423 
   1424 FunctionPass *llvm::createHexagonPacketizer() {
   1425   return new HexagonPacketizer();
   1426 }
   1427 
   1428